The leading hardware and software vendors have teamed up to build from the ground up, benchmark and actually demonstrate to the rest of the world how blazingly fast and at the same time super-efficient hyperconverged appliances can be. Intel, Mellanox, StarWind, and SuperMicro (Vendors are listed in alphabetical order), delivered highly-available (HA) compute & storage cluster featuring the latest and greatest motherboard, processor, memory, flash storage, and networking; all powered by the best hypervisor, software-defined storage (SDS), and management stacks.
Hyperconvergence is no doubt an important shift in data center technology. By moving away from the proprietary uber-expensive storage arrays to pretty much open compute & storage architectures built-out of industry-standard quite affordable x86-64 servers, extremely high bandwidth & ultra-low latency 100 Gb Ethernet interconnect fabric, proven hard disk drives for ice cold, and cutting-edge flash memory for red hot data; organizations, especially small ones, can implement the latest private cloud technologies easier and without breaking the bank. Some of the leading IT industry vendors (Intel, Mellanox, StarWind, and SuperMicro) joined their efforts to deliver a modern hyperconverged compute & storage platform to SMBs, ROBOs, and Enterprises; and demonstrate to competitors, customers and community how Hyper-Converged Infrastructure (HCI) should perform.
Intel dominates the IT landscape since mid-1970s and has absolutely no plans in stopping doing that anytime soon. It’s part of the DNA of the virtually any IT and technological innovation around the world. Intel Platinum CPUs were designed to make cloud computing happen, handle sophisticated real-time analytics, and run processing for mission-critical business applications. Intel’s Optane NVMe flash is nothing but a world’s fastest storage available today.
Mellanox Technologies is a leading provider of end-to-end, high performance InfiniBand and Ethernet interconnect fabric solutions for private and public clouds.
Mellanox pioneered RDMA (Remote Direct Memory Access) approach which is a true game changer when it comes to hyperconvergence. Their network gear provides exceptionally high transfer rates & lowest possible latency for the most public clouds enabling today's Enterprises meeting the demands of their data explosion. In a nutshell: If it’s good for Azure, it’s good for you.
StarWind. Mission-critical workloads require superlative level of uptime and breakthrough performance, at the same time nobody likes to leave money on the table. The approach of buying more storage, number crunching horsepower and so on simply doesn’t work. StarWind avoids the limitations of traditional hyperconverged infrastructure that drive up costs and cause resource underutilization by unlocking all of the underlying hardware potential. We focus on storage and our goal is to make sure 100 percent of raw IOPS are delivered within a reasonable amount of latency. No other hyperconverged vendor can do that. Period.
SuperMicro is contemporary Linux equivalent when it comes to server hardware. They are the avant-garde of IT, eager to implement all of the most recent innovations within their products. SuperMicro did a lot of heavy weight lifting to make our demo happen. We appreciate their efforts, hard work, and dedication to the project.
Our hypervisor choice is Microsoft Hyper-V and we really believe we own an explanation to the whole community here. KVM is faster, more efficient, and is where all the new development happens. VMware vSphere is a de facto standard of Enterprise grade virtualization. These two leave Hyper-V somewhere in between. So why Hyper-V over ESXi and KVM? Well… The reason we didn’t go KVM route is exactly because all the new development happens there! See, there’s simply nothing for us to code: Whole NVMe-oF stack including both target and initiator is there, it’s pretty mature at this point and there’s virtually no space for improvements, except maybe for management. Our second-best bet would be VMware vSphere, because it has no NVMe-oF target or initiator, but unfortunately VMware refused to cooperate on this issue with StarWind. Given this all, Hyper-V became the only option. In addition to that, we’ve always been a Windows shop, we really enjoy developing software for this platform and even our CEO contributed by writing some noticeable part of NVMe-oF initiator himself!
Microsoft could not provide us with a working SDS solution, capable of fully unleashing the potential of Intel’s next generation flash memory technology. When using Microsoft’s latest Storage Spaces Direct (S2D) we simply failed to obtain an expected amount of IOPS our NVMe drives could actually digest. That’s why we decided to discard S2D from our further tests and stick with proven StarWind Virtual SAN (VSAN) as software-defined storage solution of choice. Unlike S2D StarWind VSAN isn’t software architecture bound, but is limited with memory bandwidth, CPU horsepower, and network latency only.
Management didn’t go too far from the data mover. We’re disbelievers in WAC (Windows Admin Console) as it can’t perform even pretty basic tasks without dropping to PowerShell CLI. Obviously, this isn’t how GUI should work. We replaced it with our proprietary implementation, StarWind Command Center that gets the job done right.
We preserved Microsoft DiskSPD and VMFleet as they turn out to be the most trusted storage benchmarks for Microsoft Windows Server & Hyper-V. We decided to add almighty Linux FIO just to provide an alternative view on the scores.
Three benchmark scenarios are going to take place. We’ll build three slightly different test bed setups sharing most of the hardware and interconnection scheme, but using very distinct software components and settings.
The very first one is iSCSI / iSER cache-less all-flash hyperconverged cluster built as a traditional 2-node StarWind Hyper-Converged Appliance (HCA) “on steroids”, scaled out to the whooping 12 nodes this time. Each cluster node is based on the fantastic SuperMicro SuperServer 2029UZ-TR4+ motherboard, powered by a pair of just revealed to the public Intel® Xeon® Platinum 8268 processors; dual Intel® Optane™ SSD DC P4800X NVMe flash drives are responsible for storage capacity, and a pair of Mellanox ConnectX-5 100 GbE NICs handle network connectivity. Our test virtual machines will run directly from Intel Optane NVMe drives, pooled by StarWind Virtual SAN and chopped into virtual volumes. Like it was already mentioned, no cache would be used. Watch out for I/O latency!
During our second run, we’ll keep our hardware and software setup (Motherboards, CPUs, number of nodes, Hyper-V, StarWind VSAN, iSCSI / iSER, virtual volumes etc.) from above, but we’ll reconfigure our cluster to run workload from some comparably slow M.2 SATA flash drives; we’ll double the amount of Intel Optane cards, and we’ll use them as a write-back cache. At this time, we’re hunting for both IOPS are bandwidth.
And for the last but not the least, we’ll take our Hyper-Converged Infrastructure (HCI) cluster, we’ll scrap iSCSI / iSER to replace it with SPDK NVMe-oF target and StarWind NVMe-oF initiator, just to see is it possible to get even better performance than we had before. Intel Optane, NVMe-oF, no cache, the shortest possible I/O path, Windows… Exciting times ahead!
This groundbreaking demo reflects the collaboration of industry-leading IT companies featuring production-ready StarWind Hyper-Converged Appliance (HCA) built with the newest SuperMicro motherboards, the most recent Intel CPUs & NVMe flash memory, and Mellanox RDMA networking; to benchmark and expose the true Hyper-Converged Infrastructure (HCI) performance.
