Scale Computing has run lab tests showing hyperconverged virtual machine IO performance using NVM express (NVMe), achieving mean IO latencies as low as 20μs delivered to a guest virtual machine.
Scale sells HC1000, HC2000, HC4000 and HC5000 series hyperconverged infrastructure appliances (HCIA) with a HyperCore operating system and SCRIBE software storage engine.
Scale says guest virtual machines may communicate directly with SCRIBE (using standard Virtio compatible drivers) without involving a hypervisor in any way. This reduces CPU bottlenecks, context switches and system calls, all of which would add latency to each IO.
The company has been checking out the use of NVMe flash and 3D XPoint drives and announced the lab test results at VMworld, which is a bit cheeky since it uses the KVM hypervisor.
The results are good enough for it to be introducing HyperCore-Direct products, using NVMe drives, in the next few months.
Testing of HyperCore-Direct has shown that apps running with VMs can expect near bare-metal NVMe IO performance from mirrored, highly-available, software-defined storage.
An example 4-node HyperCore-Direct configuration with all-flash media is able to provide mean IO latencies of 150μs on a mixed (90 per cent read / 10 per cent write) random workload to 24 virtual machines (at a total 240,000 IOPS). With faster Opane media, the same workload achieves mean latencies of 65 microseconds.
The 4-node all-flash system has also demonstrated over 2.6m IOPS with under 290μs mean latency, as delivered to a virtualised benchmark application.
Windows 7 running IOMeter on a HyperCore-Direct system is able to obtain average latencies of 101μs for a 4k random read workload with a queue depth of 1.
Frank Ober, an Intel Optane technology solution architect, said: “Scale Computing’s solution using Intel Optane SSD P4800X is approaching the 10µs or lower bare metal latency for classic virtualisation.”
A canned quote from Howard Marks, chief scientist at DeepStorage, LLC, adds: “Scale Computing’s new HyperCore-Direct solution brings the ultra-low latency performance of NVMe to the masses. Less than a year ago, users requiring the sub-500µs latency and low overhead NVMe promised were limited to NVMe SSDs in servers or leading-edge Tier-0 systems that had no data services. Scale’s now delivering that performance on a hyperconverged platform with snapshots, and the other services enterprise applications require.”