[DRBD-user] Looking past the no-disk-{barrier,flushes} wonderland

[Sebastian Kayser]

Hi,

first thing many people seem to do when they see drbd performance issues
is to use no-disk-barrier + no-disk-flushes. I am trying to understand
some implications around this topic of ordered writes.

1) Given the above settings, drbd falls back to use WO_drain_io.
   For disks with a non-battery-backed, volatile cache this might
   increase the chance for ordered writes, but doesn't actually
   guarantee them in case of a power outage, right?

2) Is there a plan of action to very likely demo what can go wrong
   in such a setup? e.g. put a specific stress pattern on the device,
   pull the plug, reboot, and gaze in awe at a pile of corrupted data?

3) What exactly triggers ordered writes? Are they used for something
   else apart from closing epochs? Do they happen on the secondary
   only? Trying to get a feeling for the source and frequency of
   barriers/flushes.

4) Do certain access patterns in userland increase the frequency of
   drbd ordered writes? Can userland consumers or drbd be tweaked to
   ease this frequency (and thereby lessen the performance impact)
   instead of going with the sledge hammer approach of
   no-disk-barrier + no-disk-flushes?

Hope someone can shed some light.
   
Sebastian