Yellow Bricks

by Duncan Epping

VMware

How Virtual SAN enables IndonesianCloud to remain competitive!

· ·

Last week I had the chance to catch up with one of our Virtual SAN customers. I connected to Neil Cresswell through twitter and after going back and forth we got on a conference call. Neil showed me what they had created for the company he works for, a public cloud provider called IndonesianCloud. No need to tell you where they are located as the name kind of reveals it. Neil is the CEO of IndonesianCloud by the way, and very very passionate about IT / Technology and VMware. It was great talking to him, and before I forget I want to say thanks for taking time out of your busy schedule Neil, I very much appreciate it!

IndonesianCloud is a 3 year old, cloud service provider, part of the vCloud Air Network, which focuses on the delivery of enterprise class hosting services to their customers. Their customers primarily run mission critical workloads in IndonesianCloud’s three DC environment, which means that stability, reliability and predictability is really important.

Having operated a “traditional” environment for a long time Neil and his team felt it was time for a change (Servers + Legacy Storage). They needed something which was much more fit for purpose, was robust / reliable and was capable of providing capacity as well as great performance. On top of that, from a cost perspective it needed to be significantly cheaper. The traditional environment they were maintaining just wasn’t allowing them to remain competitive in their dynamic and price sensitive market. Several different hyperconverged and software based offerings were considered, but finally the settled on Virtual SAN.

Since the Virtual SAN platform was placed into production two months ago, they have deployed over 450 new virtual machines onto their initial 12 node cluster. In addition, migration of another 600 virtual machines from one of their legacy storage platforms to their Virtual SAN environment is underway. While talking to Neil I was mostly interested in some of the design considerations, some of the benefits but also potential challenges.

From a design stance Neil explained how they decided to go with SuperMicro Fat Twin hardware, 5 x NL-SAS drives (4TB) and Intel S3700 SSDs (800GB) per host. Unfortunately no affordable bigger SSDs were available, and as such the environment has a lower cache to capacity ratio than preferred. Still, when looking at the cache hit rate for reads it is more or less steady around 98-99%. PCIe flash was also looked at, but didn’t fit within the budget. These SuperMicro systems were on the VSAN Ready Node list, and this was one of the main reasons for Neil and the team to pick them. Having a pre-validated configuration, which is guaranteed to be supported by all parties, was seen as a much lower risk than building their own nodes. Then there is the network; IndonesianCloud decided to go with HP networking gear after having tested various products. One of the reasons for this was the better overall throughput, better multicast performance, and lower price per port. The network is 10GbE end to end of course.

Key take away: There can be substantial performance difference between the various 10GbE switches, do your homework!

The choice to deploy 4TB NL-SAS drives was a little risky; IndonesianCloud needed to balance the performance, capacity, and price ratios. Luckily having already run their existing cloud platform for 3 years, there was a history of IO information readily available. Using this GB/IOPS historical information meant that IndonesianCloud were able to make a calculated decision that 4TB drives with 800GB SSD would provide the perfect combination of performance and capacity. With very good cache hit rates, Neil would like to deploy larger SSD drives when they become available, as he believes that cache is a great way to minimise the impact of the slower drives. Equally, the write performance of the 4TB drives was also concerning. Using the default VSAN stripe size configuration of 1 meant that at most, only 2 drives were able to service write de-stage requests for a given VM, and due to the slow speed of the 4TB drives, this could have an impact on performance. To mitigate this, IndonesianCloud performed a series of internal tests that baselined different stripe sizes to get a good balance of performance. In the end a stripe size of 5 was selected, and is now being used for all workloads. This also helps in situations where reads are coming from disk by the way, great side effect. BTW, the best way to think about Stripe Size and Failures to Tolerate is like Raid 1E (mirrored stripes).

Key take away: Write performance of large NL-SAS drives is low, striping can help improving performance.

IndonesianCloud has standardised on a 12 node Virtual SAN cluster, and I asked why, given that Virtual SAN 5.5 U1 supports up to 32 nodes (64 with 6.0 even). Neil’s response was that 12 nodes is what comprises an internal “zone”, and that customers can balance their workloads across zones to provide higher levels of availability. Having all nodes in a single cluster, whilst possible, was not considered the best fit for a service provider that is all about containing risk. 12 nodes also maps to approximately 1000 VMs, which is what they have modelled the financial costs against, so 1000 VMs deployed on the 12 node cluster would consume CPU/Memory/Disk at the same ratio, effectively ensuring maximum utilisation of the asset.

If you look at the workloads IndonesianCloud customers run, they range from large databases, time sensitive ERP systems, webservers, streaming TV CDN services, and they are even running Airline ERP operations for a local carrier… All of these VMs are from external paying customers by the way, and all of them are mission critical for those customers. On top of Virtual SAN some customers even have other storage services running. One of them for instance is running SoftNAS on top of Virtual SAN to offer shared file services to other VMs. Vast ranges of different applications, with different IO profiles and different needs but all satisfied by Virtual SAN. One thing that Neil stressed was that the ability to change the characteristics (failures to tolerate) specified in a profile was key for them, it allows for a lot of flexibility / agility.

I did wonder, with VSAN being relative new to the market, if they had concerns in terms of stability and recoverability. Neil actually showed me their comprehensive UAT Testing Plan and the results. They were very impressed by how VSAN handled these tests without any problem. Tests ranging from pulling drives, failing network interfaces and switches, through to removing full nodes from the cluster, all of these were performed whilst simultaneously running various burn-in benchmarks. No problems whatsoever were experienced, and as a matter of fact the environment has been running great in production (don’t curse it!!).

Key take away: Testing, Testing, Testing… Until you feel comfortable with what you designed and implemented!

When it comes to monitoring though, the team did want to see more details than what is provided out of the box, especially because it is a new platform they felt that this gave them a bit more insurance that things were indeed going well and it wasn’t just their perception. They worked with one of VMware’s rock stars (Iwan Rahabok) when it comes to VR Ops on creating custom dashboards with all sorts of data ranging from cache hit ratio to latency per spindle to ANY type of detail you want on a per VM level. Of course they start with generic dashboard which then allow you to drill down; any outlier is noted immediately and leveraging VR Ops and these custom dashboards, they can drill deep whenever they need. What I loved most is how relatively easy it is for them to extend their monitoring capabilities. During our WebEx Iwan felt he needed some more specifics on a per VM basis and added these details literally within minutes to VR Ops. IndonesianCloud has been kind enough to share a custom dashboard they created, where they can catch a rogue VM easily. In this dashboard, when a single VM, and it can be any VM, generates excessive IOPS it will trigger a spike right away in the overall dashboard.

I know I am heavily biased, but I was impressed. Not just with Virtual SAN, but even more so with how IndonesianCloud has implemented it. How it is changing the way IndonesianCloud manages their virtual estate and how it enables them to compete in today’s global market.

Requirements Driven Data Center

· ·

I’ve been thinking about the term Software Defined Data Center for a while now. It is a great term “software defined” but it seems that many agree that things have been defined by software for a long time now. When talking about SDDC with customers it is typically referred to as the ability to abstract, pool and automate all aspects of an infrastructure. To me these are very important factors, but not the most important, well at least not for me as they don’t necessarily speak to the agility and flexibility a solution like this should bring. But what is an even more important aspect?

I’ve had some time to think about this lately and to me what is truly important is the ability to define requirements for a service and have the infrastructure cater to those needs. I know this sounds really fluffy, but ultimately the service doesn’t care what is running underneath, and typically the business owner and the application owners also don’t when all requirements can be met. Key is delivering a service with consistency and predictability. Even more important consistency and repeatability increase availability and predictability, and nothing is more important for the user experience.

When it comes to user experience and providing a positive one of course it is key to figure out first what you want and what you need first. Typically this information comes from your business partner and/or application owner. When you know what those requirements are then they can be translated to technical specifications and ultimately drive where the workloads end up. A good example of how this works or would look like is VMware Virtual Volumes. VVols is essentially requirements driven placement of workloads. Not just placement, but of course also all other aspects when it comes to satisfying requirements that determine user experience like QoS, availability, recoverability and whatever more is desired for your workload.

With Virtual Volumes placement of a VM (or VMDK) is based on how the policy is constructed and what is defined in it. The Storage Policy Based  Management engine gives you the flexibility to define policies anyway you like, of course it is limited to what your storage system is capable of delivering but from the vSphere platform point of view you can do what you like and make many different variations. If you specify that the object needs to thin provisioned, or has a specific IO profile, or needs to be deduplicated or… then those requirements are passed down to the storage system and the system makes its placement decisions based on that and will ensure that the demands can be met. Of course as stated earlier also requirements like QoS and availability are passed down. This could be things like latency, IOPS and how many copies of an object are needed (number of 9s resiliency). On top of that, when requirements change or when for whatever reason SLA is breached then in a requirements driven environment the infrastructure will assess and remediate to ensure requirements are met.

That is what a requirements driven solution should provide: agility, availability, consistency and predictability. Ultimately your full data center should be controlled through policies and defined by requirements. If you look at what VMware offers today, then it is fair to say that we are closing in on reaching this ideal fast.

vMSC for 6.0, any new recommendations?

· ·

I am currently updating the vSphere Metro Storage Cluster best practices white paper, over the last two weeks I received various questions if there were any new recommendation for vMSC for 6.0. I have summarized the recommendations below for your convenience, the white paper is being reviewed and I am updating screenshots, hopefully will be done soon.

  • In order to allow vSphere HA to respond to both an APD and a PDL condition vSphere HA needs to be configured in a specific way. VMware recommends enabling VM Component Protection. After the creation of the cluster VM Component Protection needs to be enabled.
  • The configuration for PDL is basic. In the “Failure conditions and VM response” section it can be configured what the response should be after a PDL condition is detected. VMware recommends setting this to “Power off and restart VMs”. When this condition is detected a VM will be restarted instantly on a healthy host within the vSphere HA cluster.
  • When an APD condition is detected a timer is started. After 140 seconds the APD condition is officially declared and the device is marked as APD time out. When the 140 seconds has passed HA will start counting, the default HA time out is 3 minutes. When the 3 minutes has passed HA will restart the impacted virtual machines, but you can configure VMCP to respond differently if desired. VMware recommends configuring it to “Power off and restart VMs (conservative)”.
    • Conservative refers to the likelihood of HA being able to restart VMs. When set to “conservative” HA will only restart the VM that is impacted by the APD if it knows another host can restart it. In the case of “aggressive” HA will try to restart the VM even if it doesn’t know the state of the other hosts, which could lead to a situation where your VM is not restarted as there is no host that has access to the datastore the VM is located on.
  • It is also good to know that if the APD is lifted and access to the storage is restored before the time-out has passed that HA will not unnecessarily restart the virtual machine, unless you explicitly configure it do so. If a response is desired even when the environment has recovered from the APD condition then “Response for APD recovery after APD timeout” should be configured to “Reset VMs”. VMware recommends leaving this setting disabled.

What does support for vMotion with active/active (a)sync mean?

· ·

Having seen so many cool features being released over the last 10 years by VMware you sometimes wonder what more they can do. It is amazing to see what level of integration we’ve see between the different datacenter components. Many of you have seen the announcements around Long Distance vMotion support by now.

When I saw this slide something stood out to me instantly and that is this part:

  • Replication Support
    • Active/Active only
      • Synchronous
      • Asynchronous

What does this mean? Well first of all “active/active” refers to “stretched storage” aka vSphere Metro Storage Cluster. So when it comes to long distance vMotion some changes have been introduced for sync stretched storage. (** note that “active/active” storage is not required for long distance vMotion**)With stretched storage writes can come from both sides at any time to a volume and will be replicated synchronously. Some optimizations have been done to the vMotion process to avoid writes during switchover to avoid any delay during the process as a result of replication traffic.

For active/active asyncronous the story is a bit different. Here again we are talking about “stretched storage” but in this case the asynchronous flavour. One important aspect which was not mentioned in the deck is that async requires Virtual Volumes. Now, at the time of writing there is no vendor yet who has a VVol capable solution that offers active/active async. But more important, is this process any different than the sync process? Yes it is!

During the migration of a virtual machine which uses virtual volumes, with an “active/active async” configuration backing it, the array is informed that a migration of the virtual machine is taking place and is requested to switch from asynchronous replication to synchronous. This to ensure that the destination is in-sync with the source when the VM is switched over from side A to side B. Besides switching from async to sync when the migration has completed the array is informed that the migration has completed. This allows the array to switch the “bias” of the VM for instance, especially in a stretched environment this is important to ensure availability.

I can’t wait for the first vendor to announce support for this awesome feature!

vVols and queueing

· ·

I was reading an article last week by Ray Lucchesi on Virtual Volumes (vVols) and queueing. In that article (and podcast) Ray (and friends on the podcast) describe vVols and the benefits they bring but also a potential danger. I have written about vVols before and if you don’t know what it is or does then I recommend reading those articles. I have been wondering as well, how all of this works, as I also felt that there could easily be a bottleneck. I had some conversations over the last couple of weeks and I figured I would share it with you instead of just leaving a comment on Ray’s blog. Lets look at an architectural diagram first:

In the diagram above (which I borrowed from the vSphere Storage blog, thanks Rolo) you see two important constructs which are part of the overall vVols architecture namely the Storage Container aka Virtual Datastore and the Protocol Endpoint (PE). The Storage Container is where the vVols will be stored. The IO though is proxied through the Protocol Endpoint. You can imagine that if we would not do this and expose every single vVol directly to vSphere that you would have 1000s of devices connected to vSphere, and as you know vSphere has a 256 device limit at the moment. This would never scale, and as such the Protocol Endpoint is used as an access point to a vVols capable storage system.

Now think about a VMFS volume and look at the vVols architectural diagram again. Yes, there is a potential bottleneck indeed. However, what the diagram does not show is that you can have multiple Protocol Endpoints. Ray mentions the following in his post: “I am also not aware of any VASA 2.0 requirement that restricts the number of PEs for a storage system’s support of a single vSphere cluster”. And I can confirm that VMware did not limit the number of Protocol Endpoints in any shape or form. I read the specifications and it literally states 1 PE at a minimum and preferably more. Note that vendor implementations of vVols may differ, I have seen implementations that describe many PEs per storage system, but also implementations which have 1 PE per storage system. And in the case of 1 PE per storage system can that be a bottleneck?

The queue depth of the Protocol Endpoint isn’t limited to 32 like a regular LUN when multiple VMs are contending for IO (“disk.schednumreqoutstanding”) or 64 (typical device queue depth) but set to 128 by default. This can be increased when required however. Before you do, please consult your storage vendor. There are a couple of variables that need to be taken in to account like the max device queue depth for instance and then there also is the HBA max queue depth as well. (For NFS queue depth is no concern typically.) The potential constraint when there is only (uncommon) a single PE can be mitigated. What is important here is that vVols itself does not impose any constraints.

Also, note that some storage vendors have an implementation where the array actually can make the distinction between regular IO and control/management related IO. Regular IO in those cases doesn’t proxy through the PE, which means you will not fill up the queue of the PE. Pretty smart.

I am hoping that clears up some of the misunderstandings out there.