The Storage Olympics
posted on August 2, 2016 by Amit Golander
No Doping Scandal Here!
Maybe a few dopes, though. After all, anyone can compete in the Storage System Olympics.
Do you have an extent-based design written entirely in a defunct minicomputer company’s command line scripting language? Bring it on. Tuned for throughput, with musclebound 200 megabyte gulps lumbering down the test rig, you might score a Gold Medal for Throughput. Even if your Frankenstein monster is useless in the real world.
Or you’ve optimized a twitchy nibbler to immediately log anything that shows up in its buffer, recording it to persistent media, but can’t keep up with it in any meaningful way. It just procrastinates, promising to actually deal with it “someday” maybe. Deferring its way to a Medal for lowest Latency. Can someone use this in a production environment characterized by constant pressure instead of burstiness, though? “Nyet!” say the Russian judges, who would score it at 1.8 or less, if they hadn’t been banned altogether from participating.
And what about parallelized teamwork? Can you push a lot of Random Accesses through your earnest little competitor? If so, are they real ops, the kind that matter to applications like databases? Or just dodgy no_op noise from some artificial benchmark nobody in their right mind should take seriously? Does it actually do the work, or just dope you by using libaio or async and delay the hard lifting to a later time in which no one watches?
Oh, like you have a better benchmark? My benchmark can eat your benchmark for breakfast, buddy. . .
All right, all right. Calm down. Let’s step back and think about what it means to be a genuine Olympian, a resident of Mount Olympus. Is narrow specialization good enough? Can you optimize for one metric and ignore everything else? At Plexistor, we don’t think so. We think a true Olympian should be a Triathlon, at the very least. In the Storage Olympics, we’d like to enter our SDM (Software-Defined Memory) software into every category of competition — and we think it will consistently win.
Can SDM outperform alternatives for low-latency and throughout and random access? In every test we’ve tried so far, that’s what we’ve seen. And that’s what we’re being told by the users around the world who are downloading it and trying it out for themselves. We’ve published a few benchmarks and case studies with performance metrics measured with FIO, and recently we’ve seen people trying these tests for themselves, using different hardware and comparing SDM to other technologies: local or across fast networks.
It’s the Storage Olympics! And we welcome you to download it (free) and test all performance parameters at once:
fio –bs=4k –rw=randrw –sync=1 –ioengine=psync –numjobs=<num_of_cpus/2> –size=<capacity/numjobs> –fallocate=none –time_based –norandommap –runtime=60 …
Benchmarking aside, we are even more pleased that our SDM makes a real difference to production environments, using its well-rounded athleticism to deliver record-breaking performance to Couchbase, Cassandra, MongoDB, MySQL and more — running on Linux OS or on Docker containers, on their own servers or on public clouds.