Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.
Hi, there is an IO request size bug in < 8.3.14 and < 8.4.2. That version model doesn't support pushing stuff to linux-stable. See this for details: http://www.gossamer-threads.com/lists/drbd/users/24104 http://git.drbd.org/gitweb.cgi?p=drbd-8.3.git;a=commit;h=3a2911f0c7ab80d061a5a6878c95f6d731c98554 Cheers, Sebastian On 24.12.2012 19:00, Digimer wrote: > On 12/24/2012 05:48 AM, Walter Robert Ditzler wrote: >> hi to all, >> >> is there anyone who has experience in speed allocation for a double 10gbit/s >> 10gbase bounded with bound type 3? >> >> i configured ... >> >> *** >> syncer { >> al-extents 31; >> rate 450M; >> *** >> >> ... but get back a copy speed of a 800mb dummy file of 46m/s. >> >> >> any help appreciated :-) >> >> happy x-mas, >> >> walter. > > First, confirm what speed your network can sustain on it's own. Use > 'iperf' if you can, it's my go-to tool for network performance testing. > Next, break out one of the backing storage devices, put a normal FS on > it (like ext4) and use something like bonnie++ (and be sure to set a > test size that is 2x your RAM) to test the bare storage maximum > sustained throughput. > > Once you know how fast they are on their own, take the slower of the two > and that is the theoretical maximum speed that DRBD will achieve. > Generally, I set the 'syncer { rate xxM; }' to ~30% of that speed. If > you set syncer to be too fast, your normal I/O will suffer as the > background syncing will consume all your available bandwidth. This is > why the sync rate is set low by default. > > If you set the 'syncer { rate xxM; }' to a speed higher than is > actually possible, it could actually hurt performance and, possibly, > stall the sync entirely. > > To be clear; > > The syncer rate is *not* the day to day performance speed. Normal I/O, > like writing a file to the FS on the DRBD resource, will always go as > fast as is possible. The syncer rate *only* effects the speed at which > out-of-sync blocks are copied over when one of the nodes has been > off-line and blocks changed on the remaining peer, > > digimer >