lkml.org 
[lkml]   [2002]   [Jun]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: 3x slower file reading oddity
dean gaudet wrote:
>
> i was trying to study various cpu & i/o bottlenecks for a backup
> tool (rdiff-backup) and i stumbled into this oddity:
>
> # time xargs -0 -n100 cat -- > /dev/null < /tmp/filelist
> 0.520u 5.310s 0:36.92 15.7% 0+0k 0+0io 11275pf+0w
> # time xargs -0 -n100 cat -- > /dev/null < /tmp/filelist
> 0.510u 5.090s 0:35.05 15.9% 0+0k 0+0io 11275pf+0w
>
> # time xargs -0 -P2 -n100 cat -- > /dev/null < /tmp/filelist
> 0.500u 5.380s 1:30.51 6.4% 0+0k 0+0io 11275pf+0w
> # time xargs -0 -P2 -n100 cat -- > /dev/null < /tmp/filelist
> 0.420u 4.810s 1:36.73 5.4% 0+0k 0+0io 11275pf+0w
>
> 3x slower with the two cats in parallel.

Note that the CPU time remained constant. The wall time went up.
You did more seeking with the dual-thread approach.

I rather depends on what is in /tmp/filelist. I assume it's
something like the output of `find'. And I assume you're
using ext2 or ext3?

- ext2/3 will chop the filesystem up into 128-megabyte block groups.

- It attemts to place all the files in a directory into the same
block group.

- It will explicitly place new directories into a different blockgroup
from their parent.

And I suspect it's the latter point which has caught you out. You have
two threads, and probably each thread's list of 100 files is from a
different directory. And hence it lives in a different block group.
And hence your two threads are competing for the disk head.

Even increasing the elevator read latency won't help you here - we don't
perform inter-file readahead, so as soon as thread 1 blocks on a read,
it has *no* reads queued up and the other thread's requests are then
serviced.

You'll get best throughput with a single read thread. There are some
smarter readahead things we could do in there, but it tends to be
that device-level readahead fixes everything up anyway.

-
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2005-03-22 13:26    [W:0.064 / U:4.040 seconds]
©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site