Tweet
Possible issue in disk benchmarking - caching.
I've just been testing a newly supplied HP E800 G4 system with (appallingly slow) NVME drive onboard and have found that the iozone tests and other benchmarks are badly gamed by memory caching.
The Hynix BC501 nvme drive is claimed (By Hewlett Packard UK's sales people) to have write speeds of 1100MB/s and reads of 1200MB/s, but in reality it's more like 145MB/s and 200MB/s - a huge difference and something I need to quantify in order to get them to replace with something that lives up to our performance requirements (I have 50 systems like this). so far when complaining about the sh*tty speeds I've been shouted down by HP, so I need something that will stand up to scrutiny.
The OS for testing is Ubuntu 18.10-beta but we deploy using RH7.5 (i8700 CPU, 32GB ram, onboard graphics card).
The iozone tests are showing write speeds for 4-8Gb files of 345MB/s and read speeds of 12,500MB/s - the latter being impossible even on NVME, so clearly cache-affected, but it appears that the write side is being substantially affected by caching too, given what I've been seeing using dd and writing directly to the drive rather than using a filesystem.
Not only do I not have confidence in these results, I clearly can't use them as proof of slowness with HP. Perhaps I'm doing this wrong, but if so I'd be grateful for some assistance.
I've just been testing a newly supplied HP E800 G4 system with (appallingly slow) NVME drive onboard and have found that the iozone tests and other benchmarks are badly gamed by memory caching.
The Hynix BC501 nvme drive is claimed (By Hewlett Packard UK's sales people) to have write speeds of 1100MB/s and reads of 1200MB/s, but in reality it's more like 145MB/s and 200MB/s - a huge difference and something I need to quantify in order to get them to replace with something that lives up to our performance requirements (I have 50 systems like this). so far when complaining about the sh*tty speeds I've been shouted down by HP, so I need something that will stand up to scrutiny.
The OS for testing is Ubuntu 18.10-beta but we deploy using RH7.5 (i8700 CPU, 32GB ram, onboard graphics card).
The iozone tests are showing write speeds for 4-8Gb files of 345MB/s and read speeds of 12,500MB/s - the latter being impossible even on NVME, so clearly cache-affected, but it appears that the write side is being substantially affected by caching too, given what I've been seeing using dd and writing directly to the drive rather than using a filesystem.
Not only do I not have confidence in these results, I clearly can't use them as proof of slowness with HP. Perhaps I'm doing this wrong, but if so I'd be grateful for some assistance.
Comment