Architecture/planning

Rethinking “storage efficiency” in HCI architectures–Part 1

Hyper-converged infrastructures (HCI) can bring several design and operational benefits to the table, adding to the long list of reasons behind its popularity. Yet, HCI also introduces new considerations in understanding and measuring technical costs associated with the architecture. These technical costs could be thought of as a usage “tax” or “overhead” on host resources. The amount attributed to this technical cost can vary quite drastically, and depends heavily on the architecture used. For an administrator, it can be a bit challenging to measure and understand. The architecture used by HCI solutions should not be overlooked, as these technical costs can not only influence the performance and consistency of the VMs, but dramatically impact the density of VMs per host, and ultimately the total cost of ownership.

With HCI, host resources (CPU, memory, and network) are now responsible for an entirely new set of duties typically provided by a storage array found in a traditional three-tier architecture. These responsibilities not only include handling VM storage I/O from end to end, but due to the distributed nature of HCI, hosts will take part in storage activity of VMs not local to the host, such as replicated writes of a VM, as well as data at rest operations and other services related to storage. These responsibilities consume host resources. The question is, how much?

This multi-part series is going to look at the basics of HCI architectures, and how they behave differently with respect to their demands on CPU, memory, and network resources. Operational comparisons are not covered in this series simply to maintain focus on the intent of this series.

"Storage efficiency" is more than what you think
The term "storage efficiency" is commonly associated with just data deduplication and compression. With hyper-converged infrastructures, this term takes on additional meaning. Storage efficiency in HCI relates to the efficiency of how I/Os are delivered to and from the VM. Efficiency of I/O delivery to and from VMs matter not only from performance and consistency as seen by the VM, but how much resource usage is introduced to the hosts in the cluster. The latter is often never considered, yet extremely important.

HCI Architectures
HCI solutions available in today’s market not only offer different data services, but are built differently, which is just one of the many reasons why it is difficult to generalize a typical amount of overhead that is needed to process storage I/O. All HCI solutions will vary (some more than others) on how they provide storage services to the VMs while maintaining resources for guest VM activity. The two basic categories, as illustrated in Figure 1 are:

Virtual appliance approach. A VM lives on each host in the cluster, delivering a distributed shared storage plane, processing I/O and the other related activities. Depending on the particular HCI solution, this virtual appliance on each host may also be responsible for a number of other duties.
Integrated/in-kernel approach. The distributed shared storage system is a part of the hypervisor, where key aspects of the storage system are part of the kernel. This allows for virtual machine I/O to traverse through the native kernel I/O path for the hosts participating in that I/O activity.

Figure 1. Comparing an I/O write between HCI architectures (simplified for clarity)

HCI solutions that use a VM to process storage I/O on each host reside in a context (user space) that is no different than application VMs running on the host. In other words, the resources allocated to this virtual appliance to perform system level storage duties, contend for the same resources as the VMs that it is trying to serve. HCI solutions built into the hypervisor maintain end-to-end control and awareness of the I/O. Since an in-kernel, integrated solution allows I/O to traverse through the native kernel I/O path, it uses the least "costly" way to use host resources. HCI solutions built into the kernel minimize the amplification of I/Os and the CPU and memory resources it takes to process those I/Os from end to end. Sometimes virtual appliance based HCI solutions will use devices on hosts configured in the hypervisor for direct pass-through (aka “VMDirectPath”) in an attempt to reduce overhead, but many of the fundamental penalties (especially as they relate to CPU cycles) of I/O amplification through this indirect path and context switching remain.

Addressing a problem in different ways
Why are their multiple approaches? Manufacturers may state many reasons why they chose a specific approach, and why their approach is superior. Most the decision comes from technical limitations and go-to-market pressures. An HCI vendor may not have the access, or the ability to provide this functionality natively in the kernel of a hypervisor. A virtual appliance approach is easier to bring to market, and naturally adaptable to different hypervisors since it is little more than a virtual machine to process storage I/O.

By way of comparison, those who have full ownership of the hypervisor can integrate this functionality directly into the hypervisor, and when appropriate, build some aspects of it right into the kernel, just as other core functionality is built into the kernel. Resource efficiency, hypervisor feature integration, as well as the contextual awareness and control of I/O types are typically the top reasons why it is beneficial to have a distributed storage mechanism built into the hypervisor.

Do both approaches work? Yes. Do both approaches produce the same result in VM behavior and host resource usage? No. Running the same workloads using HCI solutions with these two different architectures may produce very different results on the VMs, and the hosts that serve them. The degree of impact will depend on the technical cost (in resource usage) of the I/O processing, and other data services provided by a given solution.

This difference often does not show up until numerous, real workloads are put on these solutions. Just as with a traditional storage array, every solution is fast when there is little to no load on it. What counts is the behavior under real load with contending resources. This is something not always visible with synthetic testing. For HCI environments, the overall “storage efficiency” of the particular HCI solution can be better compared (assuming identical hardware and workloads) by looking at the following in a real HCI environment running production workloads:

The average number of active VMs per host when running your real workloads.
The performance characteristics of the VMs and hosts when running your real workloads while hosts are busy serving other workloads.

These measurements above take this topic from an occasionally tiresome academic debate, and demonstrates the differences in real world circumstances. Ironically, faster hardware can increase, not reduce, the differences between these architectural approaches to HCI. This is not unlike what occurs quite often now at the application level, where faster hardware exposes actual bottlenecks in software/application design previously unnoticeable with older, slower hardware.

Now that an explanation has been given as to why "storage efficiency" really means so much more than data services like deduplication and compression, the next post in this series will focus on CPU resources in HCI environments, and what to look out for when observing CPU usage behaviors in HCI environments.

Does the concept of host resource usage interest you? If so, stay tuned for the book, vSphere 6.5 Host Resources Deep Dive by Frank Denneman and Niels Hagoort. It is surely to be a must-have for those interested in the design and optimization of virtualized environments. You can also follow updates from them at @hostdeepdive on Twitter.

vSAN in cost effective independent environments

Old habits in data center design can be hard to break. New technologies are introduced that process data faster, and move data more quickly. Yet all too often, the thought process for data center design remains the same – inevitably constructed and managed in ways that reflect conventional wisdom and familiar practices. Unfortunately these common practices are often due to constraints of the technologies that preceded it, rather than aligning the current business objectives with new technologies and capabilities.

Historically, no component of an infrastructure dictated design and operation more than storage. The architecture of traditional shared storage often meant that the storage infrastructure was the oddball of the modern data center. Given enough capacity, performance, and physical ports on a fabric, a monolithic array could serve up several vSphere clusters, and therein lies the problem. The storage was not seen or treated as a clustered resource by the hypervisor like compute. This centralized way of storing data invited connectivity by as many hosts as possible in order to justify the associated costs. Unfortunately it also invited several problems. It placed limits on data center design because in part, it was far too impractical to purchase separate shared storage for every use case that would benefit from an independent environment isolated from the rest of the data center. As my colleague John Nicholson (blog/twitter) has often said, "you can’t cut your array in half." It’s a humorous, but cogent way to describe this highly common problem.

While VMware vSAN has proven to be extremely well suited for converging all applications into the same environment, business requirements may dictate a need for self contained, independent environments isolated in some manner from the rest of the data center. In "Cost Effective Independent Environments using vSAN" found on VMware’s StorageHub, I walk through four examples that show how business requirements may warrant a cluster of compute and storage dedicated for a specific purpose, and why vSAN is an ideal solution. The examples provided are:

Independent cluster management
Development/Test environments
Application driven requirements
Multi-purpose Disaster Recovery

Each example listed above details how traditional storage can fall short in delivering results efficiently, then compares how vSAN addresses and solves those specific design and operational challenges. Furthermore, learn how storage related controls are moved into the hypervisor using Storage Policy Based Management (SPBM), VMware’s framework that delivers storage performance and protection policies to VMs, and even individual VMDKs, all within vCenter. SPBM is the common management framework used in vSAN and Virtual Volumes (VVols), and is clearly becoming the way to manage software defined storage. Each example wraps up with a number of practical design tips for that specific scenario in order to get you started in building a better data center using vSAN.

Clustering is an incredibly powerful concept, and vSphere clusters in particular bring capabilities to your virtualized environment that are simply beyond comparison. With VMware vSAN, the power of clustering resources are taken to the next level, forming the next logical step in the journey of modernizing your environment in preparation for a fully software defined data center.

This use case published is the first of many more to come that are focused on practical scenarios reflecting common needs of organizations large and small, and how vSAN can help deliver results, quickly and effectively. Stay tuned!

– Pete

Accommodating for change with Virtual SAN

One of the many challenges to proper data center design is trying to accommodate for future changes, and do so in a practical way. Growth is often the reason behind change, and while that is inherently a good thing, IT budgets often don’t see that same rate of increase. CFO’s expect economies of scale to make your environment more cost efficient, and so should you.

Unfortunately, applications are always demanding more resources. The combination of commodity x86 servers and virtualization provided a flexible way to accommodate growth when it came to compute and memory resources, but addressing storage capacity and storage performance was far more difficult. Hyper-converged architectures helped break down this barrier somewhat, but some solutions lacked flexibility to cope with increasing storage capacity or performance beyond the initial prescribed configurations defined by a vendor. Users need a way to easily increase their HCI storage resources in the middle of a lifecycle without always requesting for yet another capital expenditure.

“A customer can have a car painted any color he wants as long as it’s black” — Henry Ford

But wait… it doesn’t always have to be that way. Take a look at my post on Virtual Blocks on Options in scalability with Virtual SAN. See how VSAN allows for a smarter way to approach your evolving resource needs, giving the power of choice in how you scale your environment back to you. Whether you choose to build your own servers using the VMware compatibility guide, go with VSAN Ready Nodes, or select from one of the VxRAIL options available, the principals described in the post remain the same. I hope it sparks a few ideas on how you can apply this flexibility in a strategic way to your own environment.

Thanks for reading…

How CPU related metrics in vSphere may be misinterpreted

Most Data Center Administrators are accustomed to looking for high CPU utilization rates on VMs, and the hosts in which they reside. This shouldn’t be a big surprise. After all, vCenter, and other monitoring tools have default alarms to alert against high CPU usage statistics. Features like DRS, or products that claim DRS-like functionality factor in CPU related metrics as a part of their ability to redistribute VMs under periods of contention. All of these alerts and activities suggest that high CPU values are bad, and low values are good. But what if conventional wisdom on the consumption of CPU resources is wrong?

Why should you care
Infrastructure metrics can certainly be a good leading indicator of a problem. Over the years, high CPU usage alarms have helped correctly identified many rogue processes on VMs ("Hey, who enabled the screen saver via GPO?…"). But a CPU alarm trigger assumes that high CPU usage is always bad. It also implies that the absence of an alarm condition means that there is not an issue. Both assumptions can be incorrect, which may lead to bad decision making in the Data Center.

The subtleties of performance metrics can reveal problems somewhere else in the stack – if you know how and where to look. Unfortunately, when metrics are looked at in isolation, the problems remain hidden in plain sight. This post will demonstrate how a few common metrics related to CPU utilization can be misinterpreted. Take a look at the post Observations with the Active Memory metric in vSphere to see how this can happen with other metrics as well.

The testing
There are a number of CPU related metrics to monitor in the hypervisor, and at least a couple of different ways to look at them (vCenter, and esxtop). For brevity, lets focus on two metrics that readily visible in vCenter; CPU Usage and CPU Ready. This doesn’t dismiss the importance of other CPU related metrics, or the various ways to gather them, but it is a good start to understanding the relationship between metrics. As a quick refresher, CPU Usage as it relates to vCenter has two definitions. From the host, the usage is the percentage of CPU cycles in use against the total CPU cycles available on the host. On the VM, usage shows the percent of CPU resources in use against the total available CPU cycles of the vCPUs visible to the VM. CPU Ready in vCenter measures in summation form, the amount of time that the virtual machine was ready, but could not get scheduled to run on the CPU.

A few notes about the test conditions and results:

The tests here comprise of activities that are scheduled inside each guest, and are repeated 5 times over a 1 hour period.
There are no synthetic tools used here to generate storage I/O load or consume CPU cycles. (iometer, StressLinux, etc.)
The activities performed are using processes that are only partially multithreaded. This approach is most reflective of real world environments.
The "slower" storage depicted in the testing were actually SSDs, while the "faster" storage was by leveraging PernixData FVP and distributed fault tolerant memory (DFTM) as a storage acceleration tier.
The absolute numbers are not necessarily important for this testing. The focus is more about comparing values when a variable like storage performance changes.
No shares, reservations, or limits were used on the test VMs.

The complex demands of real world environments may exhibit a much greater impact than what the testing below reveals. I reference a few actual cases of production workloads later on in the post. Synthetic load generators were not used here because they cannot properly simulate a pattern of activity that is reflective of a real environment. Synthetic load generators are good at stressing resources – not simulating real world workloads, or the time it takes for those workloads to complete their tasks.

Interpreting impacts on CPU usage and CPU Ready with changing storage performance
Looking at CPU utilization can be challenging because not all applications, nor the workloads they generate are the same. Most applications are a complex mix of some processes being multithreaded, while others are not. Some processes initiate storage I/O, while others do not. It is for this reason that we will look at CPU Usage and CPU Ready over a task that is repeated on the same sets of VMs, but using storage that performs differently.

For all practical purposes, CPU Ready doesn’t become meaningful until a host is running a large number of single vCPU VMs concurrently, or a number of multiple vCPU VMs concurrently. CPU Ready can sometimes be terribly tricky to decipher because it can be influenced in so many ways. Sometimes it may align with CPU utilization, while other times it may not. It may be affected by other resources, or it may not. It really depends on the environmental conditions. I find it a good supporting metric, but definitely not one that should stand on its own merit, without proper context of other metrics. We are measuring it here because it is generally regarded as important, and one that may contribute to load distribution activities.

Test 1: Single vCPU VM on a Host with no other activity
First let’s look at one of the very simplest of comparisons. A single vCPU VM with no other activity occurring on the host, where one test is using slower storage (blue), and the other test it is using faster storage (orange). A task was completed 5 times over the course of one hour. The image below shows that from the host perspective, peak CPU utilization increased by 79% when using the faster storage. CPU Ready demonstrated very little change, which was as expected due to the nature of this test (no other VMs running on the host).

When we look at the individual VMs, the results are similar. The images below show that CPU usage maximums for the VM increased by 24% when using the faster storage. CPU Ready demonstrated very little change here because there were no other VMs to contend with on that host. The "Storage Latency" column shows the average storage latency the VM was seeing during this time period.

You might think that higher latency may not be realistic of today’s storage technologies. The "slower" storage in this case did in fact come from SSD based storage. But remember that Flash of any kind can suffer in performance when committing larger block I/O which is quite common with real workloads. Take a look at "Understanding block sizes in a virtualized environment" for more information.

But wait… how long did the task, set to run 5 times over the period of one hour take? Well, the task took just half the time to run with the faster storage. The same amount of cycles were processing the same amount of I/Os, but just for a shorter period of time. This faster completion of a task will free up those CPU cycles for other VMs. This is the primary reason why the averages for CPU Usage and CPU Ready changed very little. Looking at this data in a timeline form in vCenter illustrates it quite clearly. There is a clear distinction of the characteristics of the task on the fast storage. Much more difficult to decipher on the run with slower storage.

Test 2: Multiple vCPU VM on a host with other activity
Now let’s let the same workload run on VMs with assigned multiple (4) vCPUs, along with other multi-vCPU VMs running in the background. This is to simulate a bit of "chatter" or activity that one might experience in a production environment.

As we can see from the images below, on the host level, both CPU usage and CPU ready values increased as storage performance increased. CPU usage maximums increased by 39% on the host. CPU Ready maximums increased by 34% on the host, which was a noticeable difference than testing without any other systems running.

When we look at the individual VMs, the results are similar. The images below show that CPU usage maximums increased by 39% with the faster storage. CPU Ready maximums increased by 51% while running on the faster storage. Considering the typical VM to host consolidation ratio, the effects can be profound.

Now let’s take a look at the timeline in vCenter to get an appreciation of how those CPU cycles were used. On the image below, you can see that like the single vCPU VM testing, the VM running on faster storage allowed for much higher CPU usage than when running on slower storage, but that it was for a much shorter period of time (about half). You will notice that in this test, the CPU Ready measurements generally increases as the CPU usage increased.

Real world examples
This all brings me back to what I witnessed years ago while administering a vSphere environment consisting of extremely CPU and storage I/O intensive workloads. Dozens of resource intensive VMs built for the purpose of compiling code. These were systems using that could multithread to near perfection – assuming storage performance was sufficient.

Now let’s look at what CPU utilization rates looked like on that same VM, running the same code compiling job where the storage environment wasn’t able to satisfy reads and writes fast enough. The same job took 46% longer to complete, all because the available CPU cycles couldn’t be used.

Still not a believer? Take a look at a presentation at the OpenStack summit by Charter Communications in April 2016, where they demonstrate exactly the effect I describe. Their Cassandra cluster deployed with VMware Integrated OpenStack, and the effects of CPU utilization when providing lower latency, higher performing storage. (key information beginning at 17:10). Their more freely breathing storage allowed CPU cycles related to storage I/O to be committed more quickly, thereby finishing the tasks much more quickly. High CPU usage was a desired result of theirs.

You might be thinking to yourself, "Won’t I have more CPU contention with faster storage?" Well, yes and no. Faster storage will give power back to the Administrator to control the usage of resources as needed, and deliver the SLAs required. And moving the point of contention to the CPU allows for what it does best; time slicing processes to complete the tasks as quickly as possible.

Sample what?
The rate at which telemetry data is sampled is a factor that can dramatically change your impression of the behavior of these resources used in the Data Center. It’s a big topic, and one that will be touched on in an upcoming post, but there is one thing to note here. When leveraging faster, lower latency storage, there are many times where CPU utilization and CPU Ready will stay the same. Why? In a real workload that involve CPU cycles executing to commit storage I/O, a workflow can may consist of a given amount of those I/Os, regardless of how long it takes. If that process took 18 seconds on slow storage, but 5 seconds on faster storage, the 20 second sampling rate within vCenter may render it in the same way. One often has to employ other tools to see these figures at a higher sampling rate. Tools such as vscsiStats and esxtop are good examples of this.

Takeaways
The testing, and examples above should make it easy to imagine a scenario in which a storage system is upgraded, and CPU related alarms are tripped more frequently, even though the processes that support a workflow have completed much more quickly. So with that, it’s good to keep the following in mind.

Slow storage will suppress CPU utilization rates – giving you the impression that from a host, or VM perspective, everything is fine.
Conversely, Fast storage will allow those CPU cycles related to storage I/O to execute, thereby increasing utilization rates – albeit for a shorter period of time. High CPU statistics are not necessarily a bad thing.
Averages and peaks can be misleading because increased utilization rates may not be recognizable in the vCenter CPU charts if it completes within the smallest sampling size (20 seconds)
Traditional methods of monitoring and balancing host resources can be misleading
Higher CPU utilization rates may not be a leading indicator of an issue. They are often be a trailing indicator of well-designed processes, or free breathing storage. Again, high CPU can be a good thing!!!
Application behavior, and the results are what counts. If a batch job in SQL takes 30 minutes, defining success should be around the desired time of that batch job. Infrastructure related metrics should help you diagnose issues and assist with achieving a desired result, but not be the one and only KPI.
Storage performance will generally impact every VM and host accessing the cluster. Whereas host based resource contention will only impact other VMs living on that same host.

Thanks for reading

– Pete

What does your infrastructure analytics really tell you?

There is no mistaking the value of data visualization combined with analytics. Data visualization can help make sense of the abstract or information not easily conveyed by numbers. Data analytics excels at taking discrete data points that make no sense on their own, into findings that have context, and relevance. The two together can present findings in a meaningful, insightful, and easy to understand way. But what are your analytics really telling you?

The problem for modern IT is that there can be an overabundance of data, with little regard to the quality of data gathered, how it relates to each other, and how to make it meaningful. All too often, this "more is better" approach obfuscates the important to such a degree that it provides less value, not more. it’s easy to collect data. The difficulty is to do something meaningful with the right data. Many tools collect metrics in an order not by which is most important, but what can be easily provided.

Various solutions with the same problem
Modern storage solutions have increased their sophistication in their analytics offerings for storage. In principle this can be a good thing, as storage capacity and performance is such a common problem with today’s environments. Storage vendors have joined the "we do that too" race of analytics features. However, feature list checkboxes can easily mask the reality – that the quality of insight is not what you might think it is. Creative license gets a little, well, creative.

Some storage solutions showcase their storage I/O analytics as a complete solution for understanding storage usage and performance of an environment. Advertising an extraordinary amount of data points collected, and sophisticated methods for collection of that data that is impressive by anyone’s standards. But these metrics are often taken at face value. Tough questions need to be asked before important decisions are made off of them. Is the right data being measured? Is the data being measure from the right location? Is the data being measured in the right way? And is the information conveyed of real value?

Accurate analytics requires that the sources of data are of the right quality and completeness. No amount of shiny presentation can override the result of using the wrong data, or using it in the wrong way.

What is the right data?
The right data has a supporting influence on the questions that you are trying to answer. Why did my application slow down after 1:18pm? How did a recent application modification impact other workloads? In Infrastructure performance, I’ve demonstrated how block sizes have historically been ignored when it came to storage design, because they could not have been easily seen or measured. Having metrics around fan speed of a storage array might be helpful for evaluating your cooling system in your Data Center, but does little to help you understand your workloads. The right data must also be collected at a rate that accurately reflects the real behavior. If your analytics offerings sample data once every 5 or 10 minutes, how can it ever show spikes of contention in resources that impact what your systems experience? The short answer is, they can’t.

The importance of location
Measuring the data at the right location is critical to accurately interpreting the conditions of your VMs, and the infrastructure in which they live. We perceive much more than we see. This is demonstrated most often with a playful optical illusion, but can be a serious problem with understanding your environment. The data gathered is often incomplete, and how you perceived it by virtue of assuming it was all the data you need all lead to the wrong conclusion. Let’s consider a common scenario where the analytics of a storage system shows great performance of a storage array, yet the VM may be performing poorly. This is the result of measuring from the wrong location. The array may have showed the latency of the components inside the device, but cannot account for latency introduced throughout the storage stack. The array metric might have been technically accurate for what it was seeing, but it was not providing you the correct, and complete metric. Since storage I/O always originate on the VMs and the infrastructure in which they live, it simply does not make sense to measure them from a supporting component like a storage array.

Measuring data inside the VM can be equally as challenging. Operating Systems’ method of data collection assume they are the sole proprietor of resources, and may not always accurately account for that fact that it is time slicing CPU clock cycles with other VMs. While the VM is the end "consumer" of resource, it also does not understand it is virtualized, and cannot see the influence of performance bottlenecks throughout the virtualization layer, or any of the physical components in the stack that support it.

VM metrics pulled from inside the guest OS may measure thing in different ways depending on Operating System. Consider the differences in how disk latency in Windows "Perfmon" is measured versus Linux "top." This is the problem with data collector based solutions that aggregate metrics from difference sources. A lot of data collected, but none of it means the same thing.

This disparate data leaves users attempting to reconcile what these metrics mean, and how they impact each other. Even worse when supposedly similar metrics from two different sources show different data. This can occur with storage array solutions that hook into vCenter to augment the array based statistics. Which one is to be believed? One over the other, or neither?

Statistics pulled solely from the hypervisor kernel avoids this nonsense. It provides a consistent method for gathering meaningful data about your VMs and the infrastructure as a whole. The hypervisor kernel is also capable of measuring this data in such a way that it accounts for all elements of the virtualization stack. However, determining the location for collection is not the end-game. We must also consider how it is analyzed.

Seeing the trees AND the forest
Metrics are just numbers. More is needed than numbers to provide a holistic understanding for an environment. Data collected that stands on its own is important, but how it contributes to the broader understanding of the environment is critical. One needs to be able to get a broad overview of an environment to drill down and identify a root cause of an issue, or be able to start out at the level of an underperforming VM and see how or why it may be impacted by others.

Many attempt to distill down this large collection of metrics to just a few that might help provide insight into performance, or potential issues. Examples of these individual metrics might include CPU utilization, Queue depths, storage latency, or storage IOPS. However, it is quite common to misinterpret these metrics when looked at in isolation.

Holistic understanding provides its greatest value when attempting to determine the impact of one workload over a group of other workloads. A VM’s transition to a new type of storage I/O pattern can often result in lower CPU activity; the exact opposite of what most would look for. The weight of impact between metrics will also vary. Think about a VM consuming large amounts of CPU. This will generally only impact other VMs on that host. In contrast, a storage based noisy neighbor can impact all VMs running on that storage system, not just the other VMs that live on that host.

Conclusion
Whether your systems are physical, virtualized, or live in the cloud, analytics exist to help answer questions, and solve problems. But analytics are far more than raw numbers. The value comes from properly digesting and correlating numbers into a story providing real intelligence. All of this is contingent on using the right data in the first place. Keep this in mind as you think about ways that you currently look at your environment.

Understanding block sizes in a virtualized environment

Cracking the mysteries of the Data Center is a bit like space exploration. You think you understand what everything is, and how it all works together, but struggle to understand where fact and speculation intersect. The topic of block sizes, as they relate to storage infrastructures is one such mystery. The term being familiar to some, but elusive enough to remain uncertain as to what it is, or why it matters.

This inconspicuous, but all too important characteristic of storage I/O has often been misunderstood (if not completely overlooked) by well-intentioned Administrators attempting to design, optimize, or troubleshoot storage performance. Much like the topic of Working Set Sizes, block sizes are not of great concern to an Administrator or Architect because of this lack of visibility and understanding. Sadly, myth turns into conventional wisdom – in not only what is typical in an environment, but how applications and storage systems behave, and how to design, optimize, and troubleshoot for such conditions.

Let’s step through this process to better understand what a block is, and why it is so important to understand it’s impact on the Data Center.

What is it?
Without diving deeper than necessary, a block is simply a chunk of data. In the context of storage I/O, it would be a unit in a data stream; a read or a write from a single I/O operation. Block size refers the payload size of a single unit. We can blame a bit of this confusion on what a block is by a bit of overlap in industry nomenclature. Commonly used terms like blocks sizes, cluster sizes, pages, latency, etc. may be used in disparate conversations, but what is being referred to, how it is measured, and by whom may often vary. Within the context of discussing file systems, storage media characteristics, hypervisors, or Operating Systems, these terms are used interchangeably, but do not have universal meaning.

Most who are responsible for Data Center design and operation know the term as an asterisk on a performance specification sheet of a storage system, or a configuration setting in a synthetic I/O generator. Performance specifications on a storage system are often the result of a synthetic test using the most favorable block size (often 4K or smaller) for an array to maximize the number of IOPS that an array can service. Synthetic I/O generators typically allow one to set this, but users often have no idea what the distribution of block sizes are across their workloads, or if it is even possibly to simulate that with synthetic I/O. The reality is that many applications draw a unique mix of block sizes at any given time, depending on the activity.

I first wrote about the impact of block sizes back in 2013 when introducing FVP into my production environment at the time. (See section "The IOPS, Throughput & Latency relationship") FVP provided a tertiary glimpse of the impact of block sizes in my environment. Countless hours with the performance graphs, and using vscsistats provided new insight about those workloads, and the environment in which they ran. However, neither tool was necessarily built for real time analysis or long term trending of block sizes for a single VM, or across the Data Center. I had always wished for an easier way.

Why does it matter?
The best way to think of block sizes is how much of a storage payload consisting in a single unit. The physics of it becomes obvious when you think about the size of a 4KB payload, versus a 256KB payload, or even a 512KB payload. Since we refer to them as a block, let’s use a square to represent their relative capacities.

Throughput is the result of IOPS, and the block size for each I/O being sent or received. It’s not just the fact that a 256KB block has 64 times the amount of data that a 4K block has, it is the amount of additional effort throughout the storage stack it takes to handle that. Whether it be bandwidth on the fabric, the protocol, or processing overhead on the HBAs, switches, or storage controllers. And let’s not forget the burden it has on the persistent media.

This variability in performance is more prominent with Flash than traditional spinning disk. Reads are relatively easy for Flash, but the methods used for writing to NAND Flash can inhibit the same performance results from reads, especially with writes using large blocks. (For more detail on the basic anatomy and behavior of Flash, take a look at Frank Denneman’s post on Flash wear leveling, garbage collection, and write amplification. Here is another primer on the basics of Flash.) A very small number of writes using large blocks can trigger all sorts of activity on the Flash devices that obstructs the effective performance from behaving as it does with smaller block I/O. This volatility in performance is a surprise to just about everyone when they first see it.

Block size can impact storage performance regardless of the type of storage architecture used. Whether it is a traditional SAN infrastructure, or a distributed storage solution used in a Hyper Converged environment, the factors, and the challenges remain. Storage systems may be optimized for different block size that may not necessarily align with your workloads. This could be the result of design assumptions of the storage system, or limits of their architecture. The abilities of storage solutions to cope with certain workload patterns varies greatly as well. The difference between a good storage system and a poor one often comes down to the abilities of it to handle large block I/O. Insight into this information should be a part of the design and operation of any environment.

The applications that generate them
What makes the topic of block sizes so interesting are the Operating Systems, the applications, and the workloads that generate them. The block sizes are often dictated by the processes of the OS and the applications that are running in them.

Unlike what many might think, there is often a wide mix of block sizes that are being used at any given time on a single VM, and it can change dramatically by the second. These changes have profound impact on the ability for the VM and the infrastructure it lives on to deliver the I/O in a timely manner. It’s not enough to know that perhaps 30% of the blocks are 64KB in size. One must understand how they are distributed over time, and how latencies or other attributes of those blocks of various sizes relate to each other. Stay tuned for future posts that dive deeper into this topic.

Traditional methods capable of visibility
The traditional methods for viewing block sizes have been limited. They provide an incomplete picture of their impact – whether it be across the Data Center, or against a single workload.

1. Kernel statistics courtesy of vscsistats. This utility is a part of ESXi, and can be executed via the command line of an ESXi host. The utility provides a summary of block sizes for a given period of time, but suffers from a few significant problems.

Not ideal for anything but a very short snippet of time, against a specific vmdk.
Cannot present data in real-time. It is essentially a post-processing tool.
Not intended to show data over time. vscsistats will show a sum total of I/O metrics for a given period of time, but it’s of a single sample period. It has no way to track this over time. One must script this to create results for more than a single period of time.
No context. It treats that workload (actually, just the VMDK) in isolation. It is missing the context necessary to properly interpret.
No way to visually understand the data. This requires the use of other tools to help visualize the data.

The result, especially at scale, is a very labor intensive exercise that is an incomplete solution. It is extremely rare that an Administrator runs through this exercise on even a single VM to understand their I/O characteristics.

2. Storage array. This would be a vendor specific "value add" feature that might present some simplified summary of data with regards to block sizes, but this too is an incomplete solution:

Not VM aware. Since most intelligence is lost the moment storage I/O leaves a host HBA, a storage array would have no idea what block sizes were associated with a VM, or what order they were delivered in.
Measuring at the wrong place. The array is simply the wrong place to measure the impact of block sizes in the first place. Think about all of the queues storage traffic must go through before the writes are committed to the storage, and reads are fetched. (It also assumes no caching tiers outside of the storage system exist). The desire would be to measure at a location that takes all of this into consideration; the hypervisor. Incidentally, this is often why an array can show great performance on the array, but suffer in the observed latency of the VM. This speaks to the importance of measuring data at the correct location.
Unknown and possibly inconsistent method of measurement. Showing any block size information is not a storage array’s primary mission, and doesn’t necessarily provide the same method of measurement as where the I/O originates (the VM, and the host it lives on). Therefore, how it is measured, and how often it is measured is generally of low importance, and not disclosed.
Dependent on the storage array. If different types of storage are used in an environment, this doesn’t provide adequate coverage for all of the workloads.

The Hypervisor is an ideal control plane to analyze the data. It focuses on the results of the VMs without being dependent on nuances of in-guest metrics or a feature of a storage solution. It is inherently the ideal position in the Data Center for proper, holistic understanding of your environment.

Eyes wide shut – Storage design mistakes from the start
The flaw with many design exercises is we assume we know what our assumptions are. Let’s consider typical inputs when it comes to storage design. This includes factors such as

Peak IOPS and Throughput.
Read/Write ratios
RAID penalties
Perhaps some physical latencies of components, if we wanted to get fancy.

Most who have designed or managed environments have gone through some variation of this exercise, followed by a little math to come up with the correct blend of disks, RAID levels, and fabric to support the desired performance. Known figures are used when they are available, and the others might be filled in with assumptions. But yet, block sizes, and everything they impact are nowhere to be found. Why? Lack of visibility, and understanding.

If we know that block sizes can dramatically impact the performance of a storage system (as will be shown in future posts) shouldn’t it be a part of any design, optimization, or troubleshooting exercise? Of course it should. Just as with working set sizes, lack of visibility doesn’t excuse lack of consideration. An infrastructure only exists because of the need to run services and applications on it. Let those applications and workloads help tell you what type of storage fits your environment best. Not the other way around.

Is there a better way?
The ideal approach for measuring the impact of block sizes will always include measuring from the location of the hypervisor, as this will provide these measurements in the right way, and from the right location. vscsiStats and vCenter related metrics are an incredible resource to tap into, and will provide the best understanding on impacts of block sizes in a storage system. There may be some time investment to decipher block size characteristics of a workload, but the payoff is generally worth the effort.

My vSphere Home Lab. 2016 edition

Here we go again. I had no intention of writing a follow-up to my "Home Lab 2015 edition" post last year, as I didn’t foresee any changes to the lab in the coming year that would be interesting enough to write about.

So much for predicting the future.

Sometimes Home Lab environments tend to border on vanity projects. I would like to think the recent changes in my lab were done out of need, but rationalizing wants into needs is common enough to be considered a national pastime. Nevertheless, my profession now has me testing workloads and new technologies on a daily basis, and this was a driving force behind these upgrades. Honest.

Demand often drives change. This is where the evolution of my Home Lab continues to mimic a production environment – just at a smaller scale. Budget, performance, capacity, space, and heat are all elements of a Home Lab design that are almost laughably similar to a production environment. Workloads evolve, and needs grow – quickly making previously used design inputs as inadequate. That is exactly what happened to me, and knew I had to invest in a few upgrades.

Compute – Performance/Testing Cluster
It was finally time to replace a few of the oldest components of the lab. My primary hosts that were built off of Intel Sandy Bridge processors used motherboards limited to just 32GB of RAM, and PCIe 2.0. I didn’t have any 10Gb connectivity without my old InfiniBand gear, and I was consistently pushing the CPUs to their limit.

I decided to go with a pair of SuperMIcro 5018D-FN4T rack mounted units. These are an incredibly small 1U form factor that feature built-in dual 10GbE and dual 1GbE interfaces, a dedicated IPMI port, a PCIe 3.0 slot, 4 drive bays, and can pack in up to 128GB of DDR4 memory. The motherboard uses the soldered on 8 core Xeon D-1540 chip and the power supply is built into the chassis. Both items reduce flexibility, but improve the no-brainer simplicity of the unit. What is most surprising when you get your hands on them is that they are incredibly small, yet still half empty when the case is cracked open. A third host will probably be in the works at some point, but it’s not necessary at this time.

It probably will come as no surprise that multiple PernixData FVP based acceleration tiers are an integral component of my infrastructure, so a few changes occurred in that realm.

1. Adding NVMe cards to use as a Flash based acceleration tier for FVP. For this lab arrangement, I used the Intel 750 NVMe based PCIe 3.0 card. While they are not officially on the VMware HCL, they are fine for the Home Lab, as they borrow heavily from the Intel DC P3xxxx line of NVMe cards that are on the VMware HCL. Intel NVMe cards are outstanding performers. Enjoy the benefits of completely bypassing all of the legacy elements the traditional storage stack on a host such as storage controllers and SCSI commands. NVMe based Flash devices is still limited by the physics of NAND Flash, but it is an incredible performer that can make any SSD based Flash drive look quite feeble in comparison. Just make sure to use Intel’s driver for vSphere.

2. More RAM to use as a DFTM acceleration tier in FVP. I placed 64GB of Micron Memory which allows me to allocate a nice chunk of RAM for FVP acceleration. The beauty of using memory as an acceleration tier avoiding all characteristics of NAND Flash, and the ability for it to leverage compression techniques. This typically increases the effective tier size between 30% and 70% depending on workload. The larger the tier size, the more content that can live in the tier, and the less eviction that occurs against the working set of data.

Compute – Management Cluster
A management cluster in a Home Lab is great. It has allowed me to really experiment with testing workloads and new technologies without any impact to the components that run the infrastructure. My Management Cluster now comprises of three Intel NUCs. I would have been perfectly happy with just a couple of NUCs as a Management cluster, but unfortunately the 16GB RAM limitation makes that a bit tough. Eventually, the NUCs will outlive their usefulness in the lab, but the great part about them is that they can easily be used as a desktop workstation, or media server. For now, they will continue to serve their purpose as a Management Cluster.

Switching
Upgrading my network meant adding 10GbE connectivity. For this, I chose a Netgear XS708E, 8 port, 10GbE switch. This would serve as a fast interconnect for east-west traffic between hosts. My adventures with InfiniBand were always interesting and educational. It’s an amazing technology, but there was just too much administrative overhead to the gear I was using. Unfortunately, there are not too many small, affordable 10GbE switches out there. The Dell 12 port X4012 10GbE switch looked really appealing based on the specs, but the ports are SFP+, so that would have meant rethinking a number of things. As for the Netgear, what do I think of it? After configuring the product, I’m convinced the folks at Netgear wanted to punish anyone who buys the unit. All of the configuration items that should be so basic in a CLI or web based UI are obfuscated in a proprietary interface that seems to be missing half of the options you’d expect. Dear Netgear, please let me configure LAGs, trunks, MTU size, and VLANs with something remotely resembling common sense. It does work, but if I could do it over, I’d choose something else.

My network core still consists of a Cisco SG300-20 Layer 3 switch. Moving away from hosts that had 6, 1GbE ports down to hosts that had just two 1GbE ports and two 10GbE ports meant that I was able to free up space on this switch. That switch still has a bit of a premium price for a 20 port L3 switch, but it has been a rock solid component of my lab for over 4 years now.

Ancillary Components
One thing I was tired of dealing with was my wireless gateway. I’ve grown sour on any consumer based WiFi/Router solutions available. Most aren’t stable, and lack features that require one to crack them with a DD-WRT build. Memory leaks and other reboot inducing behaviors are not what you want to deal with when attempting to access the lab remotely, so it was time to take a new approach. I went with the following for my gateway and wireless needs.

Motorola SB6121 DOCSIS 3.0 Cable Modem. This was purchased to replace the oversized cable modem provided by the service provider. It’s small, affordable, and prevents the cable company from changing settings on me, as they often would with their own unit.

Ubiquiti EdgeRouter PoE. This 5 port unit serves as my gateway, where one leg feeds downstream to my core switch, and another leg is used as a DMZ for my WiFi. This is a great switch that offers everything that I was looking for. Trunking, static routes, NAT and Firewalling. The multiple PoE ports makes it easy to add new wireless access points.

Ubiquiti UniFi AP Wireless Access Point. These access points pair nicely with the PoE based router above.

It’s been a rock solid, winning combination. Always on, with no random need to reboot. Total control over configuration, and no silliness from the cable provider. Mission accomplished.

Storage
This was one of the few components that didn’t change. Storage is served up by two, 5-bay Synology units with a mix of SSDs and spinning disk. I had plenty of capacity, with enough options to test various media if needed.

Mounting
Until this latest refresh, a $25 utility rack had housed the assortment of oddly shaped lab gear pretty well. With the changeover to small 1U rackmount servers and additional switchgear, it was time for an official enclosure. I went with a Tripp Lite 9U Wall Mount Cabinet. It will eventually be wall mounted, but for the time being, sits perfectly on a $12 moving dolly from Harbor Freight. The cabinet has some nice mounting ports for supplementary exhaust fans should the need arise.

Relocation
Within the first few minutes of powering up the new hosts, I realized the arrangement was going to need a new home. Server room loud? No. But moving from 38dB to 50+dB is loud enough that you wouldn’t want to be working by it all day. There is no way 1U fans spinning at 8,000 RPM will ever be soothing. I had been quite proud of how quiet my lab gear had been up until this point. I stayed away from 1U anything, and when with quiet fans wherever I could. I tried desperately to suppress the noise, replacing all of the fans with ultra-quiet Noctua fans. Unfortunately, ultra-quiet can also mean they don’t move much air. It’s not good to disregard any delta in CFM between fans. The heat alarms made it very clear this wasn’t going to work, and I didn’t want to burn up perfectly good gear. I chose to place all of the factory fans back in the 1U servers, and the 10GbE switch, and used the Noctua fans as supplementary fans in each device. They do help the primary fans to spin at a lower rate, so the effort wasn’t a total waste. The 9U cabinet will be relocated to a more permanent location than it is now, but for the time being, its making a coat closet nice and warm.

What it looks like
The entire lab, including the UPS is now self-contained, which should make its final relocation straight forward. The entire arrangement (5 hosts, 2 switches, 2 Synology NAS units, etc.) draws between 250 – 300 watts depending upon the load. Considering the old, much less capable arrangement ran at about 200 watts, I was pretty happy with the result.

In the spirit of full disclosure, the cabinet door does cover up some rather careless cable management practices. Regardless, I am thrilled with the end result and how it performs. A space efficient arrangement that is extremely powerful.

No matter how little, or how much you decide to invest in a Home Lab, I’ve learned that the satisfaction seems to be directly proportional to how much value it brings to you. Whether it be a hobby, used for professional growth, or a part of your day-to-day job duties, any sense of buyer’s remorse only seems to creep in when it’s not used. For my circumstances, that doesn’t seem to be a problem.

Working set sizes in the Data Center

There is no shortage of mysteries in the data center. These stealthy influencers can undermine performance and consistency of your environment, while remaining elusive to identify, quantify, and control. Virtualization helped expose some of this information, as it provided an ideal control plane for visibility. But it does not, and cannot properly expose all data necessary to account for these influencers. The hypervisor also has a habit of presenting the data in ways that can be misinterpreted.

One such mystery as it relates to modern day virtualized data centers is known as the "working set." This term certainly has historical meaning in the realm of computer science, but the practical definition has evolved to include other components of the Data Center; storage in particular. Many find it hard to define, let alone understand how it impacts their data center, and how to even begin measuring it.

We often focus on what we know, and what we can control. However, lack of visibility of influencing factors in the data center does not make it unimportant. Unfortunately this is how working sets are usually treated. It is often not a part of a data center design exercise because it is completely unknown. It is rarely written about for the very same reason. Ironic considering that every modern architecture deals with some concept of localization of data in order to improve performance. Cached content versus it’s persistent home. How much of it is there? How often is it accessed? All of these types of questions are critically important to know.

What is it?
For all practical purposes, a working set refers the amount of data that a process or workflow uses in a given time period. Think of it as hot, commonly accessed data of your overall persistent storage capacity. But that simple explanation leaves a handful of terms that are difficult to qualify, and quantify. What is recent? Does "amount" mean reads, writes, or both? And does it define if it is the same data written over and over again, or is it new data? Let’s explore this more.

There are a several traits of working sets that are worth reviewing.

Working sets are driven by the workload, the applications driving the workload, and the VMs that they run on. Whether the persistent storage is local, shared, or distributed, it really doesn’t matter from the perspective of how the VMs see it. The size will be largely the same.
Working sets always relate to a time period. However, it’s a continuum. And there will be cycles in the data activity over time.
Working set will comprise of reads and writes. The amount of each is important to know because reads and writes have different characteristics, and demand different things from your storage system.
Working set size refers to an amount, or capacity, but what and how many I/Os it took to make up that capacity will vary due to ever changing block sizes.
Data access type may be different. Is one block read a thousand times, or are a thousand blocks read one time? Are the writes mostly overwriting existing data, or is it new data? This is part of what makes workloads so unique.
Working set sizes evolve and change as your workloads and data center change. Like everything else, they are not static.

A simplified, visual interpretation of data activity that would define a working set, might look like below.

If a working set is always related to a period of time, then how can we ever define it? Well in fact, you can. A workload often has a period of activity followed by a period of rest. This is sometimes referred to the "duty cycle." A duty cycle might be the pattern that shows up after a day of activity on a mailbox server, an hour of batch processing on a SQL server, or 30 minutes compiling code. Taking a look over a larger period of time, duty cycles of a VM might look something like below.

Working sets can be defined at whatever time increment desired, but the goal in calculating a working set will be to capture at minimum, one or more duty cycles of each individual workload.

Why it matters
Determining a working set sizes helps you understand the behaviors of your workloads in order to better design, operate, and optimize your environment. For the same reason you pay attention to compute and memory demands, it is also important to understand storage characteristics; which includes working sets. Understanding and accurately calculating working sets can have a profound effect on the consistency of a data center. Have you ever heard about a real workload performing poorly, or inconsistently on a tiered storage array, hybrid array, or hyper-converged environment? This is because both are extremely sensitive to right sizing the caching layer. Not accurately accounting for working set sizes of the production workloads is a common reason for such issues.

Classic methods for calculation
Over the years, this mystery around working set sizes has resulted in all sorts of sad attempts at trying to calculate. Those attempts have included:

Calculate using known (but not very helpful) factors. These generally comprise of looking at some measurement of IOPS over the course of a given time period. Maybe dress it up with a few other factors to make it look neat. This is terribly flawed, as it assumes one knows all of the various block sizes for that given workload, and that block sizes for a workload are consistent over time. It also assumes all reads and writes use the same block size, which is also false.
Measure working sets defined on a storage array, as a feature of the array’s caching layer. This attempt often fails because it sits at the wrong location. It may know what blocks of data are commonly accessed, but there is no context to the VM or workload imparting the demand. Most of that intelligence about the data is lost the moment the data exits the HBA of the vSphere host. Lack of VM awareness can even make an accurately guessed cache size on an array be insufficient at times due to cache pollution from noisy neighbor VMs.
Take an incremental backup, and look at the amount of changed data. This sounds logical, but this can be misleading because it will not account for data that is written over and over, nor does it account for reads. The incremental time period of the backup may also not be representative of the duty cycle of the workload.
Guess work. You might see "recommendations" that say a certain percentage of your total storage capacity used is hot data, but this is a more formal way to admit that it’s nearly impossible to determine. Guess large enough, and the impact of being wrong will be less, but this introduces a number of technical and financial implications on data center design.

Since working sets are collected against activity that occurs on a continuum, calculating a typical working set with a high level of precision is not only impossible, but largely unnecessary. When attempting to determine working set size of a workload, the goal is to come to a number that reflects the most typical behavior of a single workload, group of workloads, or a total sum of workloads across a cluster or data center.

A future post will detail approaches that should give a sufficient level of understanding on active working set sizes, and help reduce the potential of negative impacts on data center operation due to poor guesswork.

Thanks for reading

Understanding PernixData FVP’s clustered read caching functionality

When PernixData debuted FVP back in August 2013, for me there was one innovation in particular that stood out above the rest. The ability to accelerate writes (known as “Write Back” caching) on the server side, and do so in a fault tolerant way. Leverage fast media on the server side to drive microsecond write latencies to a VM while enjoying all of the benefits of VMware clustering (vMotion, HA, DRS, etc.). Give the VM the advantage of physics by presenting a local acknowledgement of the write, but maintain all of the benefits of keeping your compute and storage layers separate.

But sometimes overlooked with this innovation is the effectiveness that comes with how FVP clusters acceleration devices to create a pool of resources for read caching (known as “Write Through” caching with FVP). For new and existing FVP users, it is good to get familiar with the basics of how to interpret the effectiveness of clustered read caching, and how to look for opportunities to improve the results of it in an environment. For those who will be trying out the upcoming FVP Freedom edition, this will also serve as an additional primer for interpreting the metrics. Announced at Virtualization Field Day 5, the Freedom Edition is a free edition of FVP with a few limitations, such as read caching only, and a maximum of 128GB tier size using RAM.

The power of read caching done the right way
Read caching alone can sometimes be perceived as a helpful way to improve performance, but temporary, and only addressing one side of the I/O dialogue. Unfortunately, this assertion tells an incomplete story. It is often criticized, but let’s remember that caching in some form is used by almost everyone, and everything. Storage arrays of all types, Hyper Converged solutions, and even DAS. Dig a little deeper, and you realize its perceived shortcomings are most often attributed to how it has been implemented. By that I mean:

Limited, non-adjustable cache sizes in arrays or Hyper Converged environments.
Limited to a single host in server side solutions. (operations like vMotion undermining its effectiveness)
Not VM or workload aware.

Existing solutions address some of these shortcomings, but fall short in addressing all three in order to deliver read caching in a truly effective way. FVP’s architecture address all three, giving you the agility to quickly adjust the performance tier while letting your centralized storage do what it does best; store data.

Since FVP allows you to choose the size of the acceleration tier, this impact alone can be profound. For instance, current NVMe based Flash cards are 2TB in size, and are expected to grow dramatically in the near future. Imagine a 10 node cluster that would have perhaps 20-40TB of an acceleration tier that may be serving up just 50TB of persistent storage. Compare this to a hybrid array that may only put in a few hundred GB of flash devices in an array serving up that same 50TB, and funneling through a pair of array controllers. Flash that the I/Os would still have traverse the network and storage stack to get to, and cached data that is arbitrarily evicted for new incoming hot blocks.

Unlike other host side caching solutions, FVP treats the collection of acceleration devices on each host as a pool. As workloads are being actively moved across hosts in the vSphere cluster, those workloads will still be able to fetch the cached content from that pool using a light weight protocol. Traditionally host based caching would have to re-warm the data from the backend storage using the entire storage stack and traditional protocols if something like a vMotion event occurred.

FVP is also VM aware. This means it understands the identity of each cached block – where it is coming from, and going to - and has many ways to maintain cache coherency (See Frank Denneman’s post Solving Cache Pollution). Traditional approaches to providing a caching tier meant that they were largely unaware of who the blocks of data were associated with. Intelligence was typically lost the moment the block exits the HBA on the host. This sets up one of the most common but often overlooked scenarios in a real environment. One or more noisy neighbor VMs can easily pollute, and force eviction of hot blocks in the cache used by other VMs. The arbitrary nature of this means potentially unpredictable performance with these traditional approaches.

How it works
The logic behind FVP’s clustered read caching approach is incredibly resilient and efficient. Cached reads for a VM can be fetched from any host participating in the cluster, which allows for a seamless leveraging of cache content regardless of where the VM lives in the cluster. Frank Denneman’s post on FVP’s remote cache access describes this in great detail.

Adjusting the charts
Since we will be looking at the FVP charts to better understand the benefit of just read caching alone, let’s create a custom view. This will allow us to really focus on read I/Os and not get them confused with any other write I/O activity occurring at the same time.

Note that when you choose a "Custom Breakdown", the same colors used to represent both reads and writes in the default "Storage Type" view will now be representing ONLY reads from their respective resource type. Something to keep in mind as you toggle between the default "Storage Type" view, and this custom view.

Looking at Offload
The goal for any well designed storage system is to deliver optimal performance to the applications. With FVP, I/Os are offloaded from the array to the acceleration tier on the server side. Read requests will be delivered to the VMs faster, reducing latency, and speeding up your applications.

From a financial investment perspective, let’s not forget the benefit of I/O “offload.” Or in other words, read requests that were satisfied from the acceleration tier. Using FVP, offload from the storage arrays serving the persistent storage tier, from the array controllers, from the fabric, and the HBAs. The more offload there is, the less work for your storage arrays and fabric, which means you can target more affordable backend storage. The hero numbers showcase the sum of this offload nicely.

Looking at Network acceleration reads
Unlike other host based solutions, FVP allows for common activities such as vMotions, DRS, and HA to work seamlessly without forcing any sort of rewarming of the cache from the backend storage. Below is an example of read I/O from 3 VMs in a production environment, and their ability to access cached reads on an acceleration device on a remote host.

Note how the Latency maintains its low latency on those read requests that came from a remote acceleration device (the green line).

How good is my read caching working?
Regardless of which write policy (Write Through or Write Back) is being used in FVP, the cache is populated in the same way.

All read requests from the backing array will place the data into the acceleration tier as it fetches it from the backing storage.
All write I/O is placed in the cache as it is written to the physical storage.

Therefore, it is easy to conclude that if read I/Os did NOT come from acceleration tier, it is from one of three reasons.

A block of data had been requested that had never been requested before.
The block of data had not been written recently, and thus, not residing in cache.
A block of data had once lived in the cache (via a read or write), but had been evicted due to cache size.

The first two items reflect the workload characteristics, while the last one is a result of a design decision – that being the cache size. With FVP you get to choose how large the devices are that make up the caching tier, so you can determine ultimately how much the solution will benefit you. Cache size can have a dramatic impact on performance because there is less pressure to evict previous data that have already been cached to make room for new data.

Visualizing the read cache usage
This is where the FVP metrics can tell the story. When looking at the "Custom Breakdown" view described earlier in this post, you can clearly see on the image below that while a sizable amount of reads were being serviced from the caching tier, the majority of reads (3,500+ IOPS sustained) in this time frame (1 week) came from the backing datastore.

Now, let’s contrast this to another environment and another workload. The image below clearly shows a large amount of data over the period of 1 day that is served from the acceleration tier. Nearly all of the read I/Os and over 60MBps of throughput that never touched the array.

When evaluating read cache sizing, this is one of the reasons why I like this particular “Custom Breakdown” view so much. Not only does it tell you how well FVP is working at offloading reads. It tells you the POTENTIAL of all reads that *could* be offloaded from the array. You get to choose how much offload occurs, because you decide on how large your tier size is, or how many VMs participate in that tier.

Hit Rate will also tell you the percentage of reads that are coming from the acceleration tier at any point and time. This can be an effective way to few cache hit frequency, but to gain more insight, I often rely on this "Custom Breakdown" to get better context of how much data is coming from the cache and backing datastores at any point in time. Eviction rate can also provide complimentary information if it shows the eviction rate creeping upward. But there can be cases were lower eviction percentages may evict enough cached data over time that it can still impact if it is still in cache. Thus the reason why this particular "Custom Breakdown" is my favorite for evaluating reads.

What might be a scenario for seeing a lot of reads coming from a backing datastore, and not from cache? Imagine running 500 VMs in an acceleration tier size of just a few GB. The working set sizes are likely much larger than the cache size, and will result in churning through the cache and not show significant demonstrable benefit. Something to keep in mind if you are trying out FVP with a very small amount of RAM as an acceleration resource. Two effective ways to make this more efficient would be to 1.) increase the cache size or 2.) decrease the number of VMs participating in acceleration. Both will achieve the same thing; providing more potential cache tier size for each VM accelerated. The idea for any caching layer is to have it large enough to hold most of the active data (aka "working set") in the tier. With FVP, you get to easily adjust the tier size, or the VMs participating in it.

Don’t know what your working set sizes are? Stay tuned for PernixData Architect!

Summary
Once you have a good plan for read caching with FVP, and arrange for a setup with maximum offload, you can drive the best performance possible from clustered read caching. On it’s own, clustered read caching implemented the way FVP does it can change the architectural discussion of how you design and spend those IT dollars. Pair this with write-buffering with the full edition of FVP, and it can change the game completely.

Dogs, Rush hour traffic, and the history of storage I/O benchmarking–Part 2

Part one of "History of storage I/O Benchmarking" attempted to demonstrate how Synthetic Benchmarks on their own simply cannot generate or measure storage performance in a meaningful way. Emulating real workloads seems like such a simple matter to solve, but in truth it is a complex problem involving technical, and non-technical challenges.

Assessing workload characteristics is often left for conjecture. Understanding the correct elements to observe is the first step to simulating them for testing, but how the problem is viewed is often left for whatever tool is readily available. Many of these tools may look at the wrong variables. A storage array might have great tools for monitoring the array, but is an incomplete view as it relates to the VM or application performance.
Understanding performance in a Datacenter crosses boundaries of subject matter expertise. A traditional Storage Administrator will see the world in the same way the array views it. Blocks, LUNS, queues, and transport protocols. Ask them about performance and be prepared for a monologue on rotational latencies, RAID striping efficiencies and read/write handling. What about the latency as seen by the VM? Don’t be surprised if that is never mentioned. It may not even be their fault, since their view of the infrastructure may be limited by access control.
When introducing a new solution that uses a technology like Flash, the word itself is seen as a superlative, not a technology. The name implies instant, fast, and other super-hero like qualities. Brilliant industry marketing, but it comes at a cost. Storage solutions are often improperly tested after some technology with Flash is introduced because conventional wisdom says it is universally faster than anything in the past. A simplified and incorrect assertion.

Evaluating performance demands a balance of understanding the infrastructure, the workloads, and the software platforms they run on. This takes time and the correct tools for insight – something most are lacking. Part one described the characteristics of real workloads that are difficult to emulate, plus the flawed approach of testing in a clustered compute environment. Unfortunately, it doesn’t end there. There is another factor to be considered; the physical characteristics of storage performance tiering layers, and the logic moving data between those layers.

Storage Performance tiering
Most Datacenters deliver storage performance using multiple persistent storage tiers and various forms of caching and buffering. Synthetic benchmarks force a behavior on these tiers that may be unrealistic. Many times this is difficult to decipher, as the tier sizes and data handling can be obfuscated by a storage vendor or unknown by the tester. What we do know is that storage tiering can certainly come in all shapes and sizes. Whether it traditional array with data progression techniques, a hybrid array, a decoupled architecture like PernixData FVP, or a Hyper Converged solution. The reality is that this tiering occurs all the time

With that in mind, there are two distinct approaches to test these environments.

Testing storage in a way to guarantee no I/O data comes from and goes to a top performing tier.
Testing storage in a way to guarantee that all I/O data comes from and goes to a top performing tier.

Which method is right for you? Both methods are neither right nor wrong as each can serve a purpose. Let’s use the car analogy again

Some might be averse to driving an electric car that only has a 100 mile range. But what if you had a commute that rarely ever went more than 30 miles a day? Think of that as like a caching/buffering tier. If a caching layer is large enough that it might serve that I/O 95% of the time, well then, it may not be necessary to focus on testing performance from that lower tier of storage.
In that same spirit, let’s say that same owner changed jobs and drove 200 miles a day. That same car is a pretty poor solution for the job. Similarly, if a storage array had just 20GB of caching/buffering for 100TB of persistent storage, the realistic working set size of each of the VMs that live on that storage would realize very little benefit from that 20GB of caching space. In that case, it would be better to test the performance of the lower tier of storage.

What about testing the storage in a way to guarantee that data comes from all tiers? Mixing a combination of the two sounds ideal, but often will not simulate the way real data will reside on the tiers, and produces a result that is difficult to determine if it reflects the way a real workload will behave. Due to the lack of identifying these caching tier sizes, or no true way to isolate a tier, this ironically ends up being the approach most commonly used – by accident alone.

When generating synthetic workloads that have a large proportion of writes, it can often be quite easy to hit buffer limit thresholds. Once again this is due to a benchmark committing every CPU cycle as a write I/O and for unrealistic periods of time. Even in extremely write intensive environments, this is completely unrealistic. It is for that reason that one can create a behavior with a synthetic benchmark against a tiered storage solution that rarely, if ever, happens in a real world environment.

When generating read I/O based synthetic tests using a large test file, those reads may sometimes hit the caching tier, and other times hit the slowest tier, which may show sporadic results. The reaction to this result often leads to running the test longer. The problem however is the testing approach, not the length of the test. Understanding the working set size of a VM is key, and should dictate how best to test in your environment. How do you determine a working set size? Let’s save that for a future post. Ultimately it is real workloads that matter, so the more you can emulate the real workloads, the better.

Storage caching population and eviction. Not all caching is the same
Caching layers in storage solutions can come in all shapes and sizes, but they depend on rules of engagement that may be difficult to determine. An example of two very important characteristics would be:

How they place data in cache. Is some sort of predictive "data progression" algorithm being used? Are the tiers using Write-Through caching to populate the cache in addition to population from data fetched from the backend storage.
How they choose to evict data from cache. Does the tier use "First-in-First-Out" (FIFO), Least Recently Used (LRU), Least Frequently Used (LFU) or some other approach for eviction?

Synthetic benchmarks do not accommodate this well. Real world workloads will depend highly on them however, and the differences show up only in production environments.

Other testing mistakes
As if there weren’t enough ways to screw up a test, here are a few other common storage performance testing mistakes.

Not testing as close to the application level as possible. This sort of functional testing will be more in line with how the application (and OS it lives on) handles real world data.
Long test durations. Synthetic benchmarks are of little use when running an exhaustive (multi-hour) set of tests. It tells very little, and just wastes time.
Overlooking a parameter on a benchmark. Settings matter because they can yield very different results.
Misunderstanding the read/write ratios of an environment. Are you calculating your ratio by IOPS, or Throughput? This alone can lead to two very different results.
Misunderstanding of typical I/O sizes in organization for reads and writes. How are you choosing to determine what the typical I/O size is?
Testing reads and writes like two independent objectives. Real workloads do not work like this, so there is little reason to test like this.
Using a final ‘score’ provided by a benchmark. The focus should be on the behavior for the duration of the test. Especially with technologies like Flash, careful attention should be paid to side effects from garbage collection techniques and other events that cause latency spikes. Those spikes matter.

Testing organizations often are vying for a position as a testing authority, or push methods or standards that somehow eliminate the mistakes described in this blog post series. Unfortunately that is not the case, but it does not matter anyway, as it is your data, and your workloads that count.

Making good use of synthetic benchmarks
It may come across that Synthetic Benchmarks or Synthetic Load Generators are useless. That is untrue. In fact, I use them all the time. Just not the way they conventional wisdom indicates. The real benefit comes once you accept the fact that they do not simulate real workloads. Here are a few scenarios in which they are quite useful.

Steady-state load generation. This is especially useful in virtualized environments when you are trying to create load against a few systems. It can be a great way to learn and troubleshoot.
Micro-benchmarking. This is really about taking a small snippet of a workload, and attempting to emulate it for testing and evaluation. Often times the test may only be 5 to 30 seconds, but will provide a chance to capture what is needed. It’s more about generating I/O to observe behavior than testing absolute performance. Look here for a good example.
Comparing independent hardware components. This is a great way to show differences an old and new SSD.
Help provide broader insight to the bigger architectural picture.

Observe, Learn and Test
To avoid wasting time "testing" meaningless conditions, spend some time in vCenter, esxtop, and other methods to capture statistics. Learn about your existing workloads before running a benchmark. Collaborating with an internal application owner can make better use of your testing efforts. For instance, if you are looking to improve your SQL performance, create a series of tests or modify an existing batch job to run inside of SQL to establish baselines and observe behavior. Test at the busiest time and the quietest time of the day, as they both provide great data points. This approach was incredibly helpful for me when I was optimizing an environment for code compiling.

Summary
Try not to lose sight of the fact that testing storage performance is not about testing an array. It’s about testing how your workloads behave against your storage architecture. Real applications always tell the real story. The reason why most dislike this answer is that it is difficult to repeat, and challenging to measure the right way. Testing the correct way can mean you might spend a little time better understanding the demand your applications put on your environment.

And here you thought you ran out of things to do for the day. Happy testing.