Drop the Mic
Andy Watson. February 14, 2019
With our latest SPEC SFS benchmark results, here at WekaIO we think we have made it crystal clear that our Matrix™ filesystem can outperform any alternative. Our latency is significantly lower (66% lower to be precise) and while using far less hardware our maximum throughput exceeds all other previously published results.
But wait! Did not NetApp just now publish a higher-throughput result for one variation of the SPEC SFS benchmark? Yes, they did. Using what would easily cost 6x more than WekaIO, NetApp grunted out more total aggregate IO, but with much slower response times at every point on their curve. Not only did they have to increase their hardware footprint by 50% to eke out an 8% higher number than we posted, they had to increase the number of clients by a commensurate amount too. And by now it ought to be obvious that this is a game of “if you throw more hardware at the benchmark then you will get more performance”. WekaIO can run on more hardware to generate more throughput, too.
The difference is that we can do that without limit, until the benchmark fails. But NetApp can only scale its cluster to 24 nodes (they are currently at 12 nodes and 72 clients), and that will only put additional stress on their benchmark response times, with latencies inching up as the cluster grows in size. Contrast that with WekaIO, where our cluster can scale linearly, and we are nowhere near any limit on the architecture. Theoretically, our cluster could grow to 32,768 nodes although we currently only support a mere 4,096. For our most recent submission we only configured 23 nodes and we think our benchmark was limited by the number of clients used to generate the load (19) rather than the storage system itself.
The benchmark highlights the difference between our scale-out architecture and NetApp’s clustered scale-up architecture. The game is over for them at 24 nodes. Their performance per client is only 29% of what WekaIO can drive per client. So, if you care about compute time utilization, NetApp isn’t going to help you.
|Number of Software Builds||6200||5700|
|Number of Clients||72||19|
|Number of Builds/Client||86||300|
When comparing to NetApp’s result we are only looking at one of the five workloads available for SPEC SFS benchmarking. WekaIO published results for all five, as you can see in the table below. In every workload WekaIO’s latency is very low. Extremely low. So low that only by caching all the Writes to DRAM (not a feasible approach to production environments!) can any other submitted result get lower. Maybe that’s why NetApp didn’t publish any of those other workloads?
|Benchmark||#1 Position||Score||ORT (ms)||#2 Position||Score||ORT (ms)|
The bottom line here is that WekaIO has proven its point, establishing that it provides the world’s fastest filesystem software and you can run it cost-effectively on industry standard hardware. High performance aside, the other appealing qualities associated with WekaIO’s Matrix (linearity with exabyte scalability, integration with GPU-accelerated workloads, thousands of zero-impact snapshots, extraordinary data durability and reliability, and simplicity which drives remarkable ease of use) are ultimately the compelling reasons for its rapidly growing adoption — not just its benchmarking prowess.