Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.
On Thu, Jan 05, 2012 at 12:06:14PM +0100, Volker wrote: > Hi, > > > Once the host is live again, i will report if that did the trick :-) > > As promised, here comes the follow-up. > > Unfortunately 8.3.12 does not do the trick. The described behaviour with > the load rising after using dd is still present. > > But i had the chance to test the I/O-Performance while the whole > environment (9 Servers having the drbd-device mounted via nfs) was under > very little use. Doing the dd's at 5am in the morning showed almost no > problems with I/O-Performance. I was even able to write 400MB to the > drbd-device without any problems regarding io-wait. > > Doing the same dd at 9am made the load go up to around 15. > > I can conclude, that since the behaviour is the same wth 8.3.8-1 and > 8.3.12, this is most likely not a drbd-bug. Having no problems under low > usage in contrast to having problems under heavier usage shows, that the > problem is the underlying I/O-Subsystem not being able handle the amount > of I/O-Requests generated by the whole environment. > > Im not sure where to go from here. If we find a solution, i'll let you > know... :-) On the server, use io-scheduler: deadline you may need to increase the number of nfsd threads. There are a few other sysfs and sysctl knobs to tune, both server and client side, to help even out write bursts and reduce latency. -- : Lars Ellenberg : LINBIT | Your Way to High Availability : DRBD/HA support and consulting http://www.linbit.com DRBD® and LINBIT® are registered trademarks of LINBIT, Austria.