Breaking Down Memory Walls

The interesting part about this presentation (for someone like me who isn't familiar with Blosc) is the chart for the AMD EPYC processor showing faster overall performance with compressed data. This is a surprisingly rare result to actually see in a realistic benchmark. You need both very fast decompression, and a (commensurately) large number of cores.

The more common tradeoff is that compression allows you to keep more data in RAM, which gives you a speedup compared to keeping (say) 1/4 the data in RAM and 3/4 on disk. But in the luxurious case that you sufficient RAM to fit all your data, it's usually faster to skip the compression and accept the memory bottleneck.

While the theory of getting faster results with compressed data is easy, this is one of the first real world examples that shows it actually happening (albeit on only one of the processors the tested). More common for benchmarks is the case shown in the full PDF, which shows that on easily compressible synthetic data the compression is an easier (though still not easy) win.

Nice work!

TLDR: testing on-the-fly memory (de)compression on contagious memory benchmarks