On (Mon) Apr 20 2009 [14:15:14], Mike Galbraith wrote: > On Mon, 2009-04-20 at 11:26 +0200, Jan Kara wrote: > > On Mon 20-04-09 11:07:17, Mike Galbraith wrote: > > > All journal modes seem subject to bad throughput under heavy pressure, > > > though data=ordered seems much less likely to suffer for some reason. > > > Major difference _seems_ to be that write()+largefile induces very much > > > swap activity. > > > My rough guess is that this depends on the VM writeout behavior. In > > ordered mode, we forcibly writeout all the dirty data on a transaction > > commit which happens every 5 seconds so they don't accumulate that much. > > Aha. > > > In other journaling modes we don't do that and decisions about writeout > > (probably how much pdflush manages to write in background vs. how much > > VM throttles the process to do the writeback itself) cause variances in > > the run time. But this is just a guess. You could gather blktraces of > > slow and fast runs and then look if the amount of IO done by different > > processes significantly differs. If Chris has merged by improvements to > > Seekwatcher, then you could nicely visualize this (hmm, that doesn't seem > > to be the case so I'm attaching the diff and a helper script - see comments > > in the beginning of the script and command helps for usage). > > Not necessary methinks. Actually _reading_ the proggy instead of just > glancing at it... > > @@ -178,7 +178,7 @@ int do_write_chunks(char *target, int *f > unsigned long long remain = len; > char *zeros; > > - zeros = calloc(1, len); > + zeros = calloc(1, data); > > I assume that was a booboo. Ouch! thanks for spotting that! I'll re-run the tests later tonight.. Amit -- To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
