Mat Keep’s blog post on InnoDB-vs-MyISAM benchmarks that Oracle recently published prompted me to do some mathematical modeling of InnoDB’s scalability as the number of cores in the server increases. Vadim runs lots of benchmarks that measure what happens under increasing concurrency while holding the hardware constant, but not as many with varying numbers of cores, so I decided to use Mat Keep’s data for this. The modeling I performed is Universal Scalability Law modeling, which can predict both software and hardware scalability, depending on how it is used.
In brief, the benchmarks are sysbench’s read-only and read-write tests, and the server has two Intel SSDs, 64GB of memory, and 4 x 12-core AMD Opteron 6172 “Magny-Cours” 2.1GHz CPUs. It is a reasonably typical commodity machine except for the high core count, which is more than I can remember seeing in the wild. The database was MySQL 5.5.7-rc. I am not sure why they didn’t run the GA version of MySQL for this benchmark. Maybe they wrote the paper before 5.5 went GA.
The following are plots of the read-only and read-write scalability models that I generated, based on these benchmarks.
The model predicts that the server will continue to provide more throughput as the core count climbs into the mid-50s, although the bang for the buck isn’t very good at that point. Also, there appears to be some bottleneck that hits more sharply than the model predicts at high core counts. It would be great if the benchmark were re-run with the same core counts and with sysbench on another machine, instead of taking 12 cores away from MySQL and giving them to sysbench. That way we could test with 48 cores and see what happens. My gut feeling is that the results will not be as good as the model predicts at high numbers of cores. But as Neil Gunther says, this wouldn’t mean the model is broken; it would mean that there is potentially something to fix in the server at high core counts. Without the model, there wouldn’t even be a basis for discussion.
The biggest thing I want to point out here is the dramatic improvement over just a few years ago, when you could “upgrade” from 4 to 8 cores and see a reduction in throughput. Oracle (and Percona, and lots of others) have done great work in the last couple of years making InnoDB scale and perform better on modern hardware.