0

I've benchmarked about 10 diffrent ssd devices with sysbench oltp_write_only.lua and I found no coroleation whatsoever between max sustained write IOPS of device (both from specification and from fio --bs=4k --iodepth=64 benchmark). I've tested nvme ssds with sustained write IOPS of 90k but those were much slower in oltp_write_only.lua test than one particular 15k write IOPS sata ssd while similar to most other sata ssds. WHY? What makes some ssd peform better in sysbench oltp_write_only.lua test? Also why oltp_write_only.lua test does not perform significantly better when mysql datadir in on tmpfs? Why raid levels 0 (2 devices), 10 (4 devices) do not affect oltp_write_only results at all? It's madness. And no, its not a bug in specific MySQL/MariaDB version - I've tried many versions and results were consistent. And yes, devices were trimmed before each test.

ndd
  • 139
  • 5
  • Might be useful to see the full fio command you used... – Anon Dec 09 '20 at 22:15
  • fio --randrepeat=1 --ioengine=libaio --direct=1 --gtod_reduce=1 --name=test --filename=test --bs=4k --iodepth=64 --size=100G --readwrite=randwrite Results from fio are close to ssd specifications. Problems are unpredictable sysbench oltp_write_only results which do not corelate with fio/specification write IOPS at all – ndd Dec 10 '20 at 15:33
  • Mmm, your fio command does look fine and doesn't have any of the usual pitfalls (e.g. missing `direct=1`). You could take a look at iostat while fio is running to double check it's not lying to you... – Anon Dec 10 '20 at 17:30

0 Answers0