Manuel Bouyer <bouyer%antioche.eu.org@localhost> writes: > On Thu, Nov 05, 2015 at 11:36:35AM +0100, Torbjörn Granlund wrote: >> [...] >> >> I have Samsung 843T "enterprise" SSD disks on all these systems. I use >> "log" as a mount option (via fstab). > > did you try disabling '-o log' ? > WAPBL will do lots of disk cache flushes, which can hurt ... Or sysctl -w vfs.wapbl.flush_disk_cache=0 One of the problems with wapbl, or really more with some disks and the whole scheme, is that when a program does fsync() that results in a log commit and then results in cache flush. That's safe but painful. Of course, arguably fsync in non-wapbl should be waiting until the bits have hit disk. The problem is not really the waiting but invoking the flush, vs getting completion interrupts for all the writes. Basically it would be better to slow down the program asking for fsync instead of everyone, since some disks seem to take a long time to flush. I had problems with rdiff-backup doing fsync excessively. Of course, turning off cache flushing is unsafe. But many things are, and it's useful debugging. I have not noticed this being a problem wtih Xen, but my systems might have disks that flush quickly (just wait for completion of all pending writes and return, without any slow operations). Or your problem could be unrelated, but definitely try it.
Attachment:
signature.asc
Description: PGP signature