The Infrastruggle

Adaptive Parallel I/O for the Win

DataCore sets another benchmark.

• By Jon William Toigo
• 08/01/2016

While running the risk of being viewed as a Johnny-One-Note, I must again draw attention to another Storage Performance Council (SPC) benchmark and another breakthrough by DataCore Software and its Adaptive Parallel I/O technology.

To catch up the latecomers: Last year, the SPC validated the results of speed and latency testing conducted by DataCore Software on a piece of software technology running on a "hyper-converged" platform comprising a "commodity" 2U Lenovo server; some DRAM; a handful of Samsung SSDs and Seagate SATA drives; and DataCore's SANsymphony storage virtualization software. The system delivered just less than 500,000 SPC-1 IOPS.

The industry was still trying to digest this when, only two months later, the SPC certified the performance of the DataCore hyper-converged single node DataCore Parallel Server, which delivered 1.5 million SPC-1 IOPS, and a hyper-converged dual-node kit (two server appliances in high-availability configuration) at 1.2 million SPC-1 IOPS in 4U.

DataCore has just demonstrated 5.1 million SPC-1 IOPS in a two-node rig (not hyper-converged), featuring a response time of .28 milliseconds. That moves the company ahead of the Huawei OceanStor 18800 V3 rig at 3 million IOPS at .92 milliseconds, and Hitachi Virtual Storage Platform G1000 and HP XP7 Storage (actually an OEM of HDS VSP) at 2 million IOPS with .96 milliseconds response time each.

Plus, the DataCore platform is less than a quarter of the price of those other hardware/software offerings, resulting in a price per IOPS of $0.10 compared to, well, seven to 10 times that cost per I/O for the other products. Needless to say, SPC is hearing some very loud complaints from some of its vendor members about the "apples to oranges" comparisons of software-defined and/or hyper-converged storage solutions vs. their hardware array monoliths; but the numbers don't lie.

Ziya Aral, DataCore co-founder, cites Moore's Law and notes that we have gone from thousands of computers to billions of computers. "Storage has always been the stepchild of the computer revolution, because of its mechanical technology. Simple physics have made everything faster -- links, devices, etc. -- but now that we have a surplus of CPU and memory, we are having to prop up the back-end -- storage."

He says that CPUs "are key" to moving the slowest portion of compute forward. That led to the development of what is now called Adaptive Parallel I/O technology, which is a cumbersome name for a simple idea: parallelizing the way that RAW I/O is handed off from the CPU to the bus in the system, then optimizing STORAGE I/O by making the best of memory caches, interconnects and storage devices. Last year, Aral resuscitated algorithms he had developed for multi-processing computers and applied them to idle cores in multi-core CPUs. The result is the stellar breakthrough in RAW I/O processing that turns commodity servers into something approaching supercomputers.

Aral says the test would have validated more than 6 million IOPS, but the benchmark wasn't designed to go that high. Technically, it had to be improved to measure 5.1 million SPC-1 IOPS, which is already a new world record.

Aral insists that, while a hyper-converged infrastructure model was used for some of his company's tests, this wasn't a validation of hyper-convergence, per se. He says he could have realized the same results in different (not hyper-converged) configurations.

"What it does prove is," Aral says, "first, all of the ‘Fibre Channel [FC] is too slow and that's why you need hyper-converged storage' talk is nonsense. We achieved our results using less than 50 percent of the FC link capacity. Frankly, the entire hyper-converged versus networked storage discussion is specious. We were measuring latencies in microseconds, not milliseconds."

Second, Aral insists that his work is paving the way for a significant breakthrough in speed. "We were running under 50 percent CPU utilization. Parallel I/O is scaling with the number of logical cores, going from 18 to 24 and beyond. As the performance of systems doubles, so does RAW I/O processing speed in a parallel I/O environment."

He sees databases, especially in-memory databases, as among the first beneficiaries of all of the speed. Analytics databases, which were once the domain of exclusive firms with deep pockets, are about to be part of everyone's infrastructure, thanks to DataCore's Adaptive Parallel I/O.