Top benefits to using Intel Optane NVMe for cache drives in VxRail
Wed, 20 May 2020 14:42:17 -0000|
Read Time: 0 minutes
Performance, endurance, and all without a price jump!
There is a saying that “A picture paints a thousand words” but let me add that a “graph can make for an awesome picture”.
Last August we at VxRail worked with ESG on a technical validation paper that included, among other things, the recent addition of Intel Optane NVMe drives for the vSAN caching layer. Figure 3 in this paper is a graph showing the results of a throughput benchmark workload (more on benchmarks later). When I do customer briefings and the question of vSAN caching performance comes up, this is my go-to whiteboard sketch because on its own it paints a very clear picture about the benefit of using Optane drives – and also because it is easy to draw.
In the public and private cloud, predictability of performance is important, doubly so for any form of latency. This is where caching comes into play, rather than having to wait on a busy system, we just leave it in the write cache inbox and get an acknowledgment. The inverse is also true. Like many parents I read almost the same bedtime stories to my young kids every night, you can be sure those books remain close to hand on my bedside “read cache” table. This write and read caching greatly helps in providing performance and consistent latency. With vSAN all-flash there no longer any read cache as the flash drives at the capacity layer provide enough random read access performance… just as my collection of bedtime story books has been replaced with a Kindle full of eBooks. Back to the write cache inbox where we’ve been dropping things off – at some point, this write cache needs to be empty, and this is where the Intel Optane NVMe drives shine. Drawing the comparison back to my kids, I no longer drive to a library to drop off books. With a flick of my finger I can return, or in cache terms de-stage, books from my Kindle back to the town library - the capacity drives if you will. This is a lot less disruptive to my day-to-day life, I don’t need to schedule it, I don’t need to stop what I’m doing, and with a bit of practice I’ve been able to do this mid story Let’s look at this in actual IT terms and business benefits.
To really show off how well the Optane drives shine, we want to stress the write cache as much as possible. This is where benchmarking tools and the right knowledge of how to apply them come into play. We had ESG design and run these benchmarking workloads for us. Now let’s be clear, this test is not reflective of a real-world workload but was designed purely to stress the write cache, in particular the de-staging from cache to capacity. The workload that created my go-to whiteboard sketch was the 100% sequential 64KB workload with a 1.2TB working set per node for 75 minutes.
The graph clearly shows the benefit of the Optane drives, they keep on chugging at 2,500MB/sec of throughput the entire time without dropping a beat. What’s not to like about that! This is usually when the techie customer in the room will try to burst my bubble by pointing out the unrealistic workload that is in no way reflective of their environment, or most environments… which is true. A more real-world workload would be a simulated relational database workload with a 22KB block size, mixing random 8K and sequential 128K I/O, with 60% reads and 40% writes, and a 600GB per node working set, which is quite a mouthful and is shown in figure 5. The results there show a steady 8.4-8.8% increase in IOPS across the board and a slower rise in latency resulting in a 10.5% lower response time under 80% load.
Those of you running OLTP workloads will appreciate the graph shown in figure 6 where HammerDB was used to emulate the database activity of a typical online brokerage firm. The Optane cache drives under that workload sustained a remarkable 61% more transactions per minute (TPM) and new orders per minute (NOPM). That can result in significant business improvement for an online brokerage firm who adopts Optane drives versus one who is using NAND SSDs.
When it comes to write cache, performance is not everything, write endurance is also extremely important. The vSAN spec requires that cache drives be SSD Endurance Class C (3,650 TBW) or above, and Intel Optane beats this hands down with an over tenfold margin at 41 PBW (41,984 TBW). The Intel Optane 3D XPoint architecture allows memory cells to be individually addressed in a dense, transistor-less, stackable design. This extremely high write endurance capability has let us spec a smaller sized cache drive, which in turn lets us maintain a similar VxRail node price point, enabling you the customer to get more performance for your dollar.
What’s not to like? Typically, you get to pick any two; faster/better/cheaper. With Intel Optane drives in your VxRail you get all three; more performance and better endurance, at roughly the same cost. Wins all around!
Author: David Glynn, Sr Principal Engineer, VxRail Tech Marketing
Related Blog Posts
Update to VxRail 7.0.100 and Unleash the Performance Within It
Mon, 02 Nov 2020 15:50:28 -0000|
Read Time: 0 minutes
What could be better than faster storage? How about faster storage, more capacity, and better durability?
Last week at Dell Technologies we released VxRail 7.0.100. This release brings support for the latest versions of VMware vSphere and vSAN 7.0 Update 1. Typically, in an update release we will see a new feature or two, but VMware out did themselves and crammed not only a load of new or significantly enhanced features into this update, but also some game changing performance enhancements. As my peers at VMware already did a fantastic job of explain these features, I won’t even attempt to replicate their work – you can find links to the blogs on features that caught my attention in the reference section below. Rather, I want to draw attention to the performance gains, and ask the question: Could RAID5 with compression only be the new normal?
Don’t worry, I can already hear the cries of “Max performance needs RAID1, RAID5 has IO amplification and parity overhead, data reduction services have drawbacks”, but bear with me a little. Also, I’m not suggesting that RAID5 compression only be used for all workloads, there are some workloads that are definitely unsuitable – streams of compressed video come to mind. Rather I’m merely suggesting that after our customers go through the painless process of updating their cluster to VxRail 7.0.100 from one of our 36 previous releases in over the past two years (yes you can leap straight from 4.5.211 to 7.0.100 in a single update and yes we do handle the converging and decommissioning of the Platform Services Controller), that they check out the reduction in storage IO latency that their existing workload is putting on their VxRail cluster, and investigate what it represents – in short, more storage performance headroom.
As customers buy VxRail clusters to support production workloads, they can’t exactly load it up with a variety of benchmark workload test to see how far they can push it. But at VxRail we are fortune to have our own dedicated performance team, who have enough VxRail nodes to run a mid-sized enterprise, and access to a large library of components so that they can replicate almost any VxRail configuration we sell (and a few we don’t). So, there is data behind my outrageous suggestion, it isn’t just back of the napkin mathematics. Grab a copy of the performance team’s recent findings in their whitepaper: Harnessing the performance of Dell EMC VxRail 7.0.100: A lab based performance analysis of VxRail, and skip to figure 3. There you’ll find some very telling before and after performance latency curves with and without data reduction services for an RDBMS workload. Spoiler: 58% more peak IOPS and almost 40% lower latency, with compression this only drops to a still very significant 45% more peak IOPS with 30% lower latency. (For those of you screaming “but failure domains” check out the blog Space Efficiency Using the New “Compression only” Option where Pete Kohler explains the issue, and how it not longer exists with compression only.) But what about RAID5? Skip up to figure 1 which summarizes the across the board performance gains for IOPS and throughput, impressive, right? Now slide down to figure 2 to compare the throughput, in particular compare RAID 1 on 7.0 with RAID 5 on 7.0 U1 – the read performance is almost identical, while the gap in write performance has narrowed. Write performance on RAID5 will likely always lag RAID1 due to IO amplification, but VMware is clearly looking to narrow that gap as much as possible.
If nothing else the whitepaper should tell you that a simple hassle-free upgrade to VxRail 7.0.100 will unlock additional performance headroom on your vSAN cluster without any additional costs, and that the tradeoffs associated with RAID5 and data reduction services (compression only) are greatly reduced. There are opportunistic space savings to be had from compression only, but they require committing to a cluster wide change to unlock, which is something that should not be taken lightly. However, realizing the guaranteed 33% capacity savings of RAID5, can be unlocked per virtual machine, reverted just as easily, represents a lower risk. I opened asking the question if RAID5 with compression only could be the new normal, and I firmly believe that the data indicates that this is a viable option for many more workloads.
My peers at VMware (John Nicholson, Pete Flecha (these two are the voices and brains behind the vSpeakingPodcast – definitely worth listening to), Teodora Todorova Hristov, Pete Koehler and Cedric Rajendran) have written great and in-depth blogs about these features that caught my attention:
vSAN HCI Mesh – eliminate stranded storage by enabling VMs registered to cluster A access storage from cluster B
Shared Witness for 2-Node Deployments - reduced administration time and infrastructure costs thru one witness for up to sixty-four 2-node clusters
Enhanced vSAN File Services – adds SMB v2.1 and v3 for Windows and Mac clients. Add Kerberos authentication for existing NFS v4.1
Space Efficiency: Compression only option - For demanding workloads that cannot take advantage of deduplication. Compression only has higher throughput, lower latency, and significantly reduced impact on write performance compared to deduplication and compression. Compression only has a reduced failure domain and 7x faster data rebuild rates.
Spare Capacity Management – Slack space guidance of 25% has been replaced with a calculated Operational Reserve the requires less space, and decreases with scale. Additional option to enable Host rebuild reserve, VxRail Sizing Tool reserves this by default when sizing configurations, with the filter Add HA
Enhanced Durability during Maintenance Mode – data being intended for a host in maintenance mode is temporally recorded in a delta file on another host, providing the configured FTT during Maintenance Mode operations
VxRail & Intel Optane for Extreme Performance
Fri, 07 Aug 2020 15:33:49 -0000|
Read Time: 0 minutes
Enabling high performance for HCI workloads is exactly what happens when VxRail is configured with Intel Optane Persistent Memory (PMem). Optane PMem provides compute and storage performance to better serve applications and business-critical workloads. So, what is Intel Optane Persistent Memory? Persistent memory is memory that can be used as storage, providing RAM-like performance, very low latency and high bandwidth. It’s great for applications that require or consume large amounts of memory like SAP HANA, and has many other use cases as shown in Figure 1 and VxRail is certified for SAP HANA as well as Intel Optane PMem.
Moreover, PMem can be used as block storage where data can be written persistently, a great example is for DBMS log files. A key advantage to using this technology is that you can start small with a single PMem card (or module), then scale and grow as needed with the ability to add up to 12 cards. Customers can take advantage of PMem immediately because there’s no need to make major hardware or configuration changes, nor budget for a large capital expenditure.
There are a wide variety of use cases today including those you see here:
Figure 1: Intel Optane PMem Use Cases
PMem offers two very different operating modes, that being Memory and App Direct, and in turn App Direct can be used in two very different ways.
First, Intel Optane PMem in Memory mode is not yet supported by VxRail. This mode acts as volatile system memory and provides significantly lower cost per GB then traditional DRAM DIMMs. A follow-on update to this blog will describe this mode and test results in much more detail once it is supported.
As for App Direct mode (supported today), PMem is consumed by virtual machines as either a block storage device, known as vPMemDisk, or as byte addressable memory, known as Virtual NVDIMM. Both provide great benefit to the applications running in a virtual machine, just in very different ways. vPMemDisk can be used by any virtual machine hardware, and by any Guest OS. Since it’s presented as a block device it will be treated like any other virtual disk. Applications and/or data can then be placed on this virtual disk. The second consumption method, NVDIMM has the advantage of being addressed in the same way as regular RAM, however, it can retain its data through reboots or power failures. This is a considerable plus for large in-memory databases like SAP HANA where cache warm-up or the time to load tables in memory can be significant!
However, it’s important to note that, like any other memory module, the PMem module does not provide data redundancy. This may not be an issue for some data files on commonly used applications that can be re-created in case of a host failure. But a key principle when using PMem, either as block storage or byte addressable memory is that the applications are responsible for handling data replication to provide durability.
New data redundancy options are expected on applications that are using PMem and should be well understood before deployment.
First, we’ll look at test results using PMem as virtual disk (or vPMemDisk). Our Engineering team tested VxRail with PMem in App Direct mode and ran comparison tests against a VxRail all-flash (P570F series platform). The testing simulated a typical 4K OLTP workload with 70/30 RW ratio. Our results achieve more than 1.8M IOPs or 6X more than the all-flash VxRail system. That equates to 93% faster response times (or lower latency) and 6X greater throughput as shown here:
Figure 2: VxRail PMem App Direct versus VxRail all-flash
This latency difference indicates the potential to improve the performance of legacy applications by placing specific data files on a PMem module, for example, placing log files on PMem. To verify the benefit of this log acceleration use case we ran a TPC-C benchmark comparing VxRail configured with a log file on a vPMEMDIsk to a VxRail all-flash vSAN, and we saw a 46% improvement on the number of transactions per minute.
Figure 3: Log file acceleration use case
For the second consumption method, we tested PMem in App direct mode using the NVDIMM consumption method. We performed tests using 1,2,4,8 and then 12 PMEM modules. All testing has been evaluated and validated by ESG (Enterprise Strategy Group). The certified white paper has been published as highlighted in the resources section.
Figure 4: NVDIMM device testing (vSAN not-optimized versus optimized PMem NVDIMM)
The results prove linear scalability as we increase the number of modules from 1 to 12. And with 12 PMem modules, VxRail achieves 80 times more IOPs than when running against vSAN not optimized (meaning VxRail all-flash vSAN with no PMem involved), and 100X for the 4K RW workload. The right half of the graphic depicts throughput results for very large IO, 64KB. When PMem is optimized on 12 modules we saw 28X higher throughput for a 64KB random read (RR) workload, and PMem is 13 times faster for the 64K RW.
What you see here is amazing performance on a single VxRail host and almost linear scalability when adding PMem!! Yes, that warrants a double bang. If you were to max out a 64-node cluster, the potential scalability is phenomenal and game changing!
So, what does all this mean? Key takeaways are:
- The local performance of VxRail with Intel Optane PMem can scale to 12M read IOPS, and more than 4M write IOPs or 70GB/s read throughput / 22GB/s write throughput on a single host.
- The use of PMEM modules doesn’t affect the regular activity on vSAN Datastores and extends the value of your VxRail platform in many ways;
- It can be used to accelerate legacy applications, such as RDBMS Log acceleration
- It enables the deployment of in memory databases and applications that can benefit from the higher IO throughput provided by PMEM while still taking the benefit of vSAN characteristics in the VxRail platform
- The local performance of a single host with 12 x 128GB PMem modules achieves more than 12M read IOPS, and more than 4M write IOPs
- It not only increases performance of traditional HCI workloads such as VDI, but also support performance-intensive transactional and analytics workloads
- It offers orders-of-magnitude faster performance than traditional storage
- It provides more memory for less cost as PMem is much less costly than DRAM
The references and validation testing have been completed by ESG (Enterprise Strategy Group). White papers and other resources on VxRail for Extreme Performance are available via the links listed below.
By: KJ Bedard – VxRail Technical Marketing Engineer