On Friday, February 8, 2002, at 04:50 , Andrew Morton wrote: > Neil Brown found that his NFS server was getting bitten by this - > pauses of a few seconds every thirty seconds, I think. His workaround > was to drastically shorten the kupdate interval so that kupdate > performs the nice smooth checkpointing for us. This is discussed > in Daniel Robbins' second article, at > > http://www-106.ibm.com/developerworks/linux/library/l-fs8/ > > ext3's checkpointing code is too drastic at present - too much, too > late, too infrequent. We do need a native solution, rather than > borrowing kupdate functionality, I guess. This was a real winner. Here's the results of some benchmarking I was doing. The sample size was 1 for each test, so don't go publishing this in any journals. ;) Copying 192MB file: NFS Ver. NFS opt. data= bdflush time 2 async ordered defaults 0:30 2 sync ordered defaults 5:46 2 sync journal defaults 4:50 3 sync journal defaults 3:10 3 sync journal Robbins 1:40 The NFS buffer size was 8K with v.2 and 32K with v.3. The journal size was 400MB. 100Mbps ethernet. Each step above was a small win, all together they add up to the sysadmins not getting lynched by the users, who had come to expect NFS async performance (before we noticed the data integrity issues). Thanks to all, -Bill P.S. just noticed the bad unintended pun above, sorry.
