Об истинной производительности

[amarao_san]

Сейчас в рассылке ceph'а кто-то начал бухтеть по поводу того, что 530 интелы фиговее 520, слово за слово, и там всплыла крайне любопытная вещь, которую я раньше не знал

Comments

[wizzard0]

bs=64k+full sync guarantees gives you 8x write amplification because native ssd block size is typically 512kb.

this also may explain the difference in performance.

[amarao_san]

Nope, the maximal size of the data chunk, available for transfer in single operaion, noted in /sys/block/*/queue/max_hw_sectors_kb.

In my case it is 32767 (kb!)

How device is processing large data chunk for writing - is separate issue. I'm not sure if there is some kind of transaction...

[wizzard0]

*Minimum* erasable unit in most SSDs is 512kb.

So, I am pretty sure that a full synchronous write will request a 512kb chunk from the SSD, read it, erase in full, append your 64k, write.

Rinse, repeat.