Le 13/06/2015 12:38, Michael van Elst a écrit :
gdt%ir.bbn.com@localhost (Greg Troxel) writes:FWIW, my preferred tool for watching this is "systat vmstat" which has not only IOPS and MB/s but % busy.iostat -x gives you separate numbers for reading and writing.
Like Greg I ignored this read/write cycle, which is confirmed by iostat. FWIW I have 128s per SU (which makes 64k stripes).
I have wd0 reads for about 3MiB/s, and wd2 reads for 3MiBs/s then gets written for 3MiB/s.
I think that the slowness is I/O disk bound there: both disks were used elsewhere first, so both have quite different block content. Given that these disks are rather slow (5400 RPM ones), I suppose that they dislike 64k read => compare => then 64k write on one, hence the parity rebuild takes quite some time.
Good to know; that means that I should have zero'ed both disks first so as to avoid the 64k write on one of the disk which tends to cripple the whole rebuild process due to the R/W cycle.
The peak at the beginning was likely due to the zero'ing I did initially on both disks to start with fresh MBR and labels. That made the write step unnecessary, hence the faster rebuild for the first minute or so.
Interesting. Thanks! -- Jean-Yves Migeon