Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.
Christian Christian Hammers wrote: >On Thu, Jan 22, 2004 at 11:23:43AM +0100, Francis SOUYRI wrote: > > >>>But the fact that I got much better values with dd and rsync *while* >>>the syncall was in progress does not speak for a great influence. >>> >>> >>> >>dd and rsync do this: read from disk node1, network, write to MEMORY >>node2 (buffer/cache). >>drbd syncall with "Wire protocol: C" do this: read from disk node1, >>network, write TO DISK node2. >> >>This is not equivalent. >> >> > >Well, of course, I rsync a file from /tmp on the one server to /tmp >on the other one which should be equivalent! > > Your /tmp is a tmpfs, ramfs filesystem (memory filesystem) or a disk filesystem ? If it is a disk filesystem the test of the dd/rsync and drbd syncall with a "Wire protocol: C" is not equivalent, if you want to do an equivalent test you can try a test with a "Wire protocol: B". What is the filesystem of our drbd devices what is the block size ? for the drbd device a 4k block size is better. > > >>>I attached these. >>> >>>I did a throughput test with >>> dd if=/dev/zero of=/mnt/nb0/... bs=1M count=100 >>>on the master and got the following iostat values on the primary: >>>Device: tps Blk_read/s Blk_wrtn/s Blk_read Blk_wrtn >>>dev8-0 300.00 54312.00 696.00 54312 696 >>>which is about 10x as fast as durinc syncall although it is replicated, >>>too. >>> >>> >>> >>dd write to the drbd device (of=/mnt/nb0/...) but "iostat" display lot >>of reads (Blk_read/s 54312.00) ?!? >> >> > >Oups, this was the wrong cut&paste. I just tried it again and the >numbers are at about 30000 Blk_wrtn/s on both sides. > >bye, > >-christian- > > > Bye, Francis