Nutanix AOS 5.5 delivers 1M read IOPS from a single VM, but what about 70/30 read/write?

I recently wrote Nutanix AOS 5.5 delivers 1M IOPS from a single VM, but what happens when you vMotion which showed the impact of a vMotion was around -10% for a period of approx. 3 seconds before read performance resumed back to pre-migration levels.

In this post I will be addressing the question about performance for a single VM with a more realistic 70% read, 30% write IO profile which was performed using an 8k IO size and what the impact is during and after a live migration.

While not surprising to Nutanix customers, this result shows a maximum starting baseline of 436K random read and 187k random write IOPS and immediately following the migration performance reduced to 359k read and 164k write IOPS before achieving greater performance than the original baseline @ 446k read and 192k IOPS within a few seconds.

So in comparison to 100% random read which achieved just over 1 million 8k IOPS, the 70/30 mix achieves in the ballpark of 600k IOPS which is very respectable. Not bad for a platform which Nutanix competitors continue to describe as only being good for VDI. Considering even the largest array from a leading all flash SAN vendor is only advertising performance in the hundreds of thousand random read range, it shows Nutanix unique hyper-converged architecture can achieve higher performance than a monolithic all flash array from a single VM.

This shows that with the unique Nutanix Acropolis Distributed Storage Fabric, very high performance at low latency can be achieved with real world IO patterns even during and after live migrating the virtual machine across a distributed platform.

This result is further evidence of the efficiency of Nutanix Acropolis Hypervisor, AHV (which is included at no additional charge with AOS) as well as the IO path running in user space (not the much hyped in-kernel). This is in part thanks to AHV Turbo Mode which optimised the IO path which was announced at .NEXT 2017 in Washington. In addition to these excellent levels of performance, they can be sustained even when using data protection features such as snapshots as shown in recent post I wrote about Nutanix X-ray tool where I used the Snapshot impact scenario to compare Nutanix AHV and a leading hypervisor and SDS product. If you don’t have time to read the post, in short, the Nutanix competitors performance degraded as snapshots were taken while Nutanix AHV’s performance remained consistent which is essential for real world scenarios, especially with business critical applications.

With Nutanix unique ability to scale out performance using storage only nodes, even higher performance can be achieved without modification to the virtual machine to applications which gives Nutanix further advantage over the competition.

Nutanix data locality ensures optimal performance by ensuring new data is always local to the VM and cold data can remain remote indefinitely while only hot data will be migrated locally if/when required at a 1MB granularity. This translates to intelligent data locality and not brute force locality as it is frequently mistaken to be.

Back to Part 1

What’s .NEXT 2017 – AHV Turbo Mode

Back in 2015 I wrote a series titled “Why Nutanix Acropolis Hypervisor (AHV) is the next generation hypervisor” which covered off many reasons why AHV was and would become a force to be reckoned with.

In short, AHV is the only purpose built hypervisor for hyper-converged infrastructure (HCI) and it has continued to evolve in terms of functionality and maturity while becoming a popular choice for customers.

How popular you ask? Nutanix officially reported 23% adoption as a percentage of nodes sold in our recent third quarter fiscal year 2017 financial highlights.

Over the last couple of years I have personally worked with numerous customers who have adopted AHV especially when it comes to business critical applications such as MS SQL, MS Exchange.

One such example is Shinsegae who is a major retailer running 50,000 MS Exchange mailboxes on Nutanix using AHV as the hypervisor. Shinsegae also runs MS SQL workloads on the same platform which has now become the standard platform for all workloads.

This is just one example of AHV proven in the field and at scale to have the functionality, resiliency and performance to support business critical workloads.

But at Nutanix we’re always striving to deliver more value to our customers, and one area where there is a lot of confusion and misinformation is around the efficiency of the storage I/O path for Nutanix.

The Nutanix Controller VM (CVM) runs on top of multiple hypervisors and delivers excellent performance, but there is always room for improvement. With our extensive experience with in-kernel and virtual machine based storage solutions, we quickly learned that the biggest bottleneck is the hypervisor itself.

Next2017TraditionalHypervisorBottleneck

With technology such as NVMe becoming mainstream and 3D XPoint not far behind, we looked for a way to give customers the best value from these premium storage technologies.

That’s where AHV Turbo mode comes into play.

Next2017AHVTurboMode

AHV Turbo mode is a highly optimised I/O path (shortened and widened) between the User VM (UVM) and Nutanix stargate (I/O engine).

These optimisation have been achieved by moving the I/O path in-kernel.

V

V

V

V

V

V

V

V

V

V

V

Just kidding! In-kernel being better for performance is just a myth, Nutanix has achieved major performance improvements by doing the heavy lifting of the I/O data path in User Space, which is the opposite of the much hyped “In-kernel”.

The below diagram show the UVM’s I/O path now goes via Frodo (a.k.a Turbo Mode) which runs in User Space (not In-kernel) and onto stargate within the Controller VM).

Next2017FrodoDatapath2

Another benefit of AHV and Turbo mode is that it eliminates the requirement for administrators to configure multiple PVSCSI adapters and spread virtual disks across those controllers. When adding virtual disks to an AHV virtual machine, disks automatically benefit from Nutanix SCSI and block multi-queue ensuring enhanced I/O performance for both reads and writes.

The multi-queue I/O flow is handled by multiple frodo threads (Turbo mode) threads and passed onto stargate.

Next2017FrodoUpdated

As the above diagram shows, Nutanix with Turbo mode eliminates the bottlenecks associated with legacy hypervisors, one such example is VMFS datastores which required VAAI Atomic Test and Set (ATS) to minimise the impact of locking when the numbers of VMs per datastore increased (e.g. >25). With AHV and Turbo mode, every vdisk has always had it’s own queue (not one per datastore or container) but frodo enhances this by adding a per-vcpu queue at the virtual controller level.

How much performance improvement you ask? Well I ran a quick test which showed amazing performance improvements even on a more than four year old IVB NX3450 which only has 2 x SATA SSDs per node and with the memory read cache disabled (i.e.: No reads from RAM).

A quick summary of the findings were:

  1. 25% lower CPU usage for the similar sequential write performance (2929MBps vs 2964MBps)
  2. 27.5% higher sequential read performance (9512MBps vs 7207MBps)
  3. A 62.52% increase in random read IOPS (510121 vs 261265)
  4. A 33.75% increase in random write IOPS (336326 vs 239193)

So with Turbo Mode, Nutanix is using less CPU and RAM to drive higher IOPS & throughput and doing so in user space.

Intel published “Code Sample: Hello World with Storage Performance Development Kit and NVMe Driver” which states “When comparing the SPDK userspace NVMe driver to an approach using the Linux Kernel, the overhead latency is up to 10x lower”.

This is just one of many examples which shows userspace is clearly not the bottleneck that some people/vendors have tried to claim with the “in-kernel” is faster nonsense I have previously written about.

With Turbo mode, AHV is the highest performance (throughput / IOPS) and lowest latency hypervisor supported by Nutanix!

But wait there’s more! Not only is AHV now the highest performing hypervisor, it’s also used by our largest customer who has more than 1750 nodes running 100% AHV!

Next20171650Nodes

VMware you’re full of it (FUD) : Nutanix CVM/AHV & vSphere/VSAN overheads

For a long time now, VMware & EMC are leading the charge along with other vendors spreading FUD regarding the Nutanix Controller VM (CVM), making claims it uses a lot of resources to drive storage I/O and it being a Virtual Machine (a.k.a Virtual Storage Appliance / VSA) is inefficient / slower than running In-Kernel.

An recent example of the FUD comes from an article written by the President of VCE himself, Mr Chad Sakac who wrote:

… it (VxRAIL) is the also the ONLY HCIA that has a fully integrated SDS stack that is embedded into the kernel – specifically VSAN because VxRail uses vSphere.  No crazy 8vCPU, 16+ GB of RAM for the storage stack (per “storage controller” or even per node in some cases with other HCIA choices!) needed.

So I thought I would put together a post covering what the Nutanix CVM provides and giving a comparison to what Chad referred to as a fully integrated SDS stack.

Let’s compare what resources are required between the Nutanix suite which is made up of the Acropolis Distributed Storage Fabric (ADSF) & Acropolis Hypervisor (AHV) and VMware’s suite made up of  vCenter , ESXi ,  VSAN and associated components.

This should assist those not familiar with the Nutanix platform understand the capabilities and value the CVM provides and correct the FUD being spread by some competitors.

Before we begin, let’s address the default size for the Nutanix CVM.

As it stands today, The CVM by default is assigned 8 vCPUs and 16GB RAM.

What CPU resources the CVM actually uses obviously depends on the customers use case/s so if the I/O requirements are low, the CVM wont use 8 vCPU, or even 4vCPUs, but it is assigned 8vCPUs.

With the improvement in ESXi CPU scheduling over the years, the impact of having more than the required vCPUs assigned to a limited number of VMs (such as the CVM) in an environment is typically negligible, but the CVM can be right sized which is also common.

The RAM allocation is recommended to be 24Gb when using deduplication, and for workloads which are very read intensive, the RAM can be increased to provide more read cache.

However, increasing the CVM RAM for read cache (Extent Cache) is more of a legacy recommendation as the Acropolis Operating System (AOS) 4.6 release achieves outstanding performance even with the read cache disabled.

In fact, the >150K 4k random read IOPS per node which AOS 4.6 achieves on the NX-9040-G4 nodes was done without the use of in-memory read cache as part of engineering testing to see how hard the SSD drives can be pushed. As a result, even for extreme levels of performance, increasing the CVM RAM for Read Cache is no longer a requirement. As such, 24Gb RAM will be more than sufficient for the vast majority of workloads and reducing RAM levels is also on the cards.

Thought: Even if it was true in-kernel solutions provided faster outright storage performance, (which is not the case as I showed here), this is only one small part of the equation. What about management? VSAN management is done via vSphere Web Client which runs in a VM in user space (i.e.: Not “In-Kernel”) which connects to vCenter which also runs as a VM in user space which commonly leverage an SQL/Oracle database which also runs in user space.

Now think about Replication, VSAN uses vSphere Replication, which, you guessed it, runs in a VM in user space. For Capacity/Performance management, VSAN leverages vRealise Operations Manager (vROM) which also runs in user space. What about backup? The vSphere Data Protection appliance is yet another service which runs in a VM in user space.

All of these products require the data to move from kernel space into user space, So for almost every function apart from basic VM I/O, VSAN is dependant on components which are running in user space (i.e.: Not In-Kernel).

Lets take a look at the requirements for VSAN itself.

According to the VSAN design and sizing guide (Page 56) VSAN uses up to 10% of hosts CPU and requires 32GB RAM for full VSAN functionality. Now the RAM required doesn’t mean VSAN is using all 32GB and the same is true for the Nutanix CVM if it doesn’t need/use all the assigned RAM, it can be downsized although 12GB is the recommended minimum, 16GB is typical and for a node with even 192Gb RAM which is small by todays standards, 16GB is <10% which is minimal overhead for either VSAN or the Nutanix CVM.

In my testing VSAN is not limited to 10% CPU usage and this can be confirmed in VMware’s own official testing of SQL in : VMware Virtual SAN™ Performance with Microsoft SQL Server

In short, the performance testing is conducted with 3 VMs each with 4 vCPUs each on hosts contained a dual-socket Intel Xeon Processor E5-2650 v2 (16 cores, 32 threads, @2.6GHz).

So assuming the VMs were at 100% utilisation, they would only be using 75% of the total cores (12 of 16).  As we can see from the graph below, the hosts were almost 100% utilized, so something other than the VMs is using the CPU. Best case, VSAN is using ~20% CPU, with the hypervisor using 5%, in reality the VMs wont be pegged at 100% so the overhead of VSAN will be higher than 20%.

VSAN10PercentBS

Now I understand I/O requires CPU, and I don’t have a problem with VSAN using 20% or even more CPU, what I have a problem with is VMware lying to customers that it only uses 10% AND spreading FUD about other vendors virtual appliances such as the Nutanix CVM are resource hogs.

Don’t take my word for it, do your own testing and read their documents like the above which simple maths shows the claim of 10% max is a myth.

So that’s roughly 4 vCPUs (on a typical dual socket 8 core system) and up to 32GB RAM required for VSAN, but lets assume just 16GB RAM on average as not all systems are scaled to 5 disk groups.

The above testing was not on the latest VSAN 6.2, so things may have changed. One such change is the introduction of software checksums into VSAN. This actually reduces performance (as you would expect) because it provides a layer of data integrity with every I/O, as such the above performance is still a fair comparison because Nutanix has always had software checksums as this is essential for any production ready storage solution.

Now keep in mind, VSAN is really only providing the storage stack, so its using ~20% CPU under heavy load for just the storage stack, unlike the Nutanix CVM which is also providing a highly available management layer which has comparable (and in many cases better functionality/availability/scalability) to vCenter, VUM, vROM, vSphere Replication, vSphere Data Protection, vSphere Web Client, Platform Services Controller (PSC) and the supporting database platform (e.g.: SQL/Oracle/Postgress).

So I comparing VSAN CPU utilization to a Nutanix CVM is about as far from Apples/Apples as you could get, so let’s look at what all the vSphere Managements components resource requirements are and make a fairer comparison.

vCenter Server

Resource Requirements:

Small | Medium | Large

  • 4vCPUs | 8vCPUs | 16vCPUs
  • 16GB | 24GB | 32GG RAM

Reference:http://pubs.vmware.com/vsphere-60/index.jsp#com.vmware.vsphere.install.doc/GUID-D2121DC5-1FC8-48DC-A4BA-C3FD72D0BE77.html

Platform Services Controller

Resource Requirements:

  • 2vCPUs
  • 2GB RAM

Reference:http://pubs.vmware.com/vsphere-60/index.jsp#com.vmware.vsphere.install.doc/GUID-D2121DC5-1FC8-48DC-A4BA-C3FD72D0BE77.html

vCenter Heartbeat (Deprecated)

If we we’re to compare apples to apples, vCenter would need to be fully distributed and highly available which its not. The now deprecated vCenter Heartbeat used to be able to somewhat provide, so that’s 2x the resources of vCenter, VUM etc, but since its deprecated we’ll give VMware the benefit of the doubt and not count resources to make their management components highly available.

What about vCenter Linked Mode? 

I couldn’t find its resource requirements in the documentation so let’s give VMware the benefit of the doubt and say it doesn’t add any overheads. But regardless of overheads, its another product to install/validate and maintain.

vSphere Web Client

The Web Client is required for full VSAN management/functionality and has its own resource requirements:

  • 4vCPUs
  • 2GB RAM (at least)

Reference:https://pubs.vmware.com/vsphere-50/index.jsp#com.vmware.vsphere.install.doc_50/GUID-67C4D2A0-10F7-4158-A249-D1B7D7B3BC99.html

vSphere Update Manager (VUM)

VUM can be installed on the vCenter server (if you are using the Windows Installation) to save having management VM and OS to manage, if you are using the Virtual Appliance then a seperate windows instance is required.

Resource Requirements:

  • 2vCPUs
  • 2GB

The Nutanix CVM provides the ability to do Major and Minor patch updates for ESXi and of course for AHV.

vRealize Operations Manager (vROM)

Nutanix provides built in Analytics similar to what vROM provides in PRISM Element and centrally managed capacity planning/management and “what if” scenarios for adding nodes to the cluster, as such including vROM in the comparison is essential if we want to get close to apples/apples.

Resource Requirements:

Small | Medium | Large

  • 4vCPUs | 8vCPUs | 16vCPUs
  • 16GB | 32GB | 48GB
  • 14GB Storage

Remote Collectors Standard | Large

  • 2vCPUs | 4vCPUs
  • 4GB | 16GB

Reference:https://pubs.vmware.com/vrealizeoperationsmanager-62/index.jsp#com.vmware.vcom.core.doc/GUID-071E3259-625A-437B-AB34-E6A58B87C65B.html

vSphere Data protection

Nutanix also has built in backup/recovery/snapshotting capabilities which include application consistency via VSS. As with vROM we need to include vSphere Data Protection in any comparison to the Nutanix CVM.

vSphere Data Protection can be deployed in 0.5 to 8TB as shown below:

VDPrequirements

Reference: http://pubs.vmware.com/vsphere-60/topic/com.vmware.ICbase/PDF/vmware-data-protection-administration-guide-61.pdf

The minimum size is 4vCPUs and 4GB RAM but that only supports 0.5TB, for even an average size node which supports say 4TB, 4 vCPUs and 8GB is required.

So best case scenario we need to deploy one VDP appliance per 8TB, which is smaller than some Nutanix (or VSAN Ready) nodes (e.g.: NX6035 / NX8035 / NX8150) so that would potentially mean one VDP appliance per node when running VSAN since the backup capabilities are not built in like they are with Nutanix.

Now what about if I want to replicate my VMs or use Site Recovery Manager (SRM)?

vSphere Replication

As with vROM and vSphere Data protection, vSphere Replication provides VSAN functionality which Nutanix also has built into the CVM. So we also need to include vSphere Replication resources in any comparison to the CVM.

While vSphere Replication has fairly light on resource requirements, if all my replication needs to go via the appliance, it means one VSAN node will be a hotspot for storage and network traffic, potentially saturating the network/node and being a noisy neighbour to any Virtual machines on the node.

Resource Requirements:

  • 2vCPUs
  • 4GB RAM
  • 14GB Storage

Limitations:

  • 1 vSphere replication appliance per vCenter
  • Limited to 2000 VMs

Reference: http://pubs.vmware.com/vsphere-replication-61/index.jsp?topic=%2Fcom.vmware.vsphere.replication-admin.doc%2FGUID-E114BAB8-F423-45D4-B029-91A5D551AC47.html

So scaling beyond 2000 VMs requires another vCenter, which means another VUM, another Heartbeat VM (if it was still available), potentially more databases on SQL or Oracle.

Nutanix doesn’t have this limitation, but again we’ll give VMware the benefit of the doubt for this comparison.

Supporting Databases

The size of even a small SQL server is typically at least 2vCPUs and 8GB+ RAM and if you want to compare apples/apples with Nutanix AHV/CVM you need to make the supporting database server/s highly available.

So even in a small environment we would be talking 2 VMs @ 2 vCPUs and 8GB+ RAM ea just to support the back end database requirements for vCenter, VUM, SRM etc.

As the environment grows so does the vCPU/vRAM and Storage (Capacity/IOPS) requirements, so keep this in mind.

So what are the approx. VSAN overheads for a small 4 node cluster?

The table below shows the minimum vCPU/vRAM requirements for the various components I have discussed previously to get VSAN comparable (not equivalent) functionality to what the Nutanix CVM provides.

VSANoverheads3

As the above only covers the minimum requirements for a small say 4 node environment, things like vSphere Data Protection will require multiple instances, SQL should be made highly available using an Always on Availability group (AAG) which requires a 2nd SQL server and as the environment grows, so do the vCPU/vRAM requirements for vCenter, vRealize Operations Manager and SQL.

A Nutanix AHV environment on the other hand looks like this:

NutanixOverheads2

So just 32 vCPUs and 64GB RAM for a 4 node cluster which is 8vCPU and 54GB RAM LESS than the comparable vSphere/VSAN 4 node solution.

If we add Nutanix Scale out File Server functionality into the mix (which is optionally enabled) this increases to 48vCPUs and 100GB RAM. Just 8vCPUs more and still 18GB RAM LESS than vSphere/VSAN while Nutanix provides MORE functionality (e.g.: Scale out File Services) and comes out of the box with a fully distributed, highly available, self healing FULLY INTEGRATED management stack.

The Nutanix vCPU count assumes all vCPUs are in use which is VERY rarely the case. So this comparison is well and truely in favour of VSAN while still showing vSphere/VSAN having higher overheads for a typical/comparable solution with Nutanix providing additional built in features such as Scale out File Server (another distributed and highly available solution) for only a small amount more resources than vSphere/VSAN which does not provide comparable native file serving functionality.

What about if you don’t use all those vSphere/VSAN features and therefore don’t deploy all those management VMs. VSAN overheads are lower, right?

It is a fair argument to say not all vSphere/VSAN features need to be deployed, so this will reduce the vSphere/VSAN requirements (or overheads).

The same however is true for the Nutanix Controller VM.

Its not uncommon where customers don’t run all features and/or have lower I/O requirements for the CVM to be downsized to 6vCPUs. I personally did this earlier this week for a customer running SQL/Exchange this week and the CVM is still only running at ~75% or approx ~4 vCPUs and that’s running vBCA with in-line compression.

So the overheads depend on the workloads, and the default sizes can be changed for both vSphere/VSAN components and the Nutanix CVM.

Now back to the whole In-Kernel nonsense.

VMware also like to spread FUD that their own hypervisor has such high overheads, its crazy to run any storage through it. I’ve always found this funny since VMware have been telling the market for years the hypervisor has a low overhead (which it does), but they change their tune like the weather to suit their latest slideware.

One such example of this FUD comes from VMware’s Chief Technologist, Duncan Epping who tweeted:

DuncanInKernelRubbish

The tweet is trying to imply that going through the hypervisor to another Virtual Machine (in this case a Nutanix CVM) is inefficient, which is interesting for a few reasons:

  1. If going from one VM to another via the kernel has such high overheads, why do VMware themselves recommend virtualizing business critical high I/O applications which have applications access data between VMs (and ESXi hosts) all the time? e.g.: When a Web Server VM accesses an Application Server VM which accesses data from a Database. All this is in one VM, through the kernel and into another VM.
  2. Because for VSAN has to do exactly this to leverage many of the features it advertises such as:
  • Replication (via vSphere Replication)
  • vRealize Operations Manager (vROM)
  • vSphere Data Protection (vDP)
  • vCenter and supporting components

Another example of FUD from VMware, in this case Principal Engineer, Jad El-Zein is implying VSAN has low(er) overheads compared to Nutanix (Blocks = Nutanix “Blocks”):

screenshot_2015-03-10-10-05-44-1.png

I guess he forgot about the large number of VMs (and resources) required to provide VSAN functionality and basic vSphere management. Any advantage of being In-Kernel (assuming you still believe it is in fact any advantage) are well and truely eliminated by the constant traffic across the hypervisor to and from the management VMs all of which are not In-Kernel as shown below.

vSphereManagement

I’d say its #AHVisTheOnlyWay and #GoNutanix since the overheads of AHV are lower than vSphere/VSAN!

Summary:

  1. The Nutanix CVM provides a fully integrated, preconfigured and highly available, self healing management stack. vSphere/VSAN requires numerous appliances and/or software to be installed.
  2. The Nutanix AHV Management stack (provided by the CVM) using just 8vCPUs and typically 16GB RAM provides functionality which in many cases exceeds the capabilities of vSphere/VSAN which requires vastly more resources and VMs/Appliances to provide comparable (but in many cases not equivalent) functionality.
  3. The Nutanix CVM provides these capabilities built in (with the exception of PRISM Central which is a seperate Virtual Appliance) rather than being dependant on multiple virtual appliances, VMs and/or 3rd party database products for various functionality.
  4. The Nutanix management stack is also more resilient/highly available that competing products such as all VMware management components and comes this way out of the box. As the cluster scales, the Acropolis management stack continues to automatically scale management capabilities to ensure linear scalability and consistent performance.
  5. Next time VMware/EMC try to spread FUD about the Nutanix Controller VM (CVM) being a resource hog or similar, ask them what resources are required for all functionality they are referring to. They probably haven’t even considered all the points we have discussed in this post so get them to review the above as a learning experience.
  6. Nutanix/AHV management is fully distributed and highly available. Ask VMware how to make all the vSphere/VSAN management components highly available and what the professional services costs will be to design/install/validate/maintain that solution.
  7. The next conversation to have would be “How much does VSAN cost compared to Nutanix”? Now that we understand all the resources overheads and complexity in design/implementation/validation of the VSAN/vSphere environment, not to mention most management components will not be highly available beyond vSphere HA. But cost is a topic for another post as the ELA / Licensing costs are the least of your worries.

To our friends at VMware/EMC, the Nutanix CVM says,

“Go ahead, underestimate me”.