Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.
On Jul 25, 2011, at 3:59 PM, listslut at outofoptions.net wrote: > Quoting Zev Weiss <zweiss at scout.wisc.edu>: > >> Hi, >> >> Ken, Mark -- any further information on the write performance problem, and/or a solution to it? >> >> I'm running 8.3.7 at the moment, I think I'm going to try updating to 8.3.11 and/or 8.4 and see if it magically fixes anything (unless there are further suggestions on other things to try). >> >> >> Zev >> > I've put together a quick test cluster. DRBD does well on the initial disk sync. I just went back and looked at the graphs to make sure it wasn't my imagination. I'm beginning to think it has something to do with the interaction of DRBD/LVM2/GFS2 that is not happy. I did most of the tweaks I could find and it added up to 50% increase, but going from 1.5M to 2.0M isn't exactly thrilling. > > Ken Lowther > Yeah, I think I've seen decent performance on sync as well, but when it comes to actual usage as the backing store of a live filesystem, things go rapidly downhill -- I've been seeing about 4MB/s on my system. Not quite the 2M you're seeing, but still pretty unacceptably slow, by an order of magnitude or so. As for the source of the problem, I don't have GFS2 in the picture on my setup, so I think that'd rule out that component (assuming we're looking at the same underlying cause). I may be able to attach some extra drives to my machines and create another drbd resource to test with, so I can try e.g. with and without LVM between DRBD and the raw block devices and see if that makes a difference. Zev (I seem to have a bad habit of mistakenly replying to the sender instead of the list...sorry about that Ken.)