[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: writing processes are blocking in log_wait_common with data=ordered

On Tue, 30 Apr 2002, Andrew Morton wrote:

> David Mansfield wrote:
> > 
> > ...
> > One interesting and unexpected result is that running inside a looped back
> > filesystem 1gb in size increases performance 4-fold from running on the
> > real filesystem!  That is, ext3-ordered looped on top of ext3-ordered is
> > much faster than ext3-ordered!  This is on kernel 2.4.17-rc2-aa2, which is
> > a bit old, so it could be meaningless....
> Heh.  The loop driver tells lies: when it signals I/O
> completion, your data is still floating about in memory,
> unwritten.  So yup, it'll run quicker.
> > P.S.  I created a benchmark of this phenomenon called blktest.c.
> Great stuff, thanks.

I've done some more testing, and there's definitely something fishy going 
on here.  I tested on my home computer (1GHZ Athlon, ATA133 IDE disk) and 
got numbers like:

Kernel 2.4.19-pre7

150 blks/sec ext2
100 blks/sec ext3-ordered
96 blks/sec ext3-writeback

Kernel 2.4.19-pre7-aa3

168 blks/sec ext2
98 blks/sec ext3-ordered
96 blks/sec ext3-writeback.

So on this system, it seems like ext2 is a 50% improvement.  Maybe this is 
just to be expected...

And for reference, on the server our application is running on (1ghz 
Intel, SCSI U160 drive, 2.4.19-pre7-aa2, 2gb ram with blk high io enabled)

44 blks/sec ext3-ordered. (half as fast as that inferior system).

The last thing I tried is running the program on a server here running the 
beta of 'Red Hat Advanced Server' which is running their kernel 
'2.4.9-26beta.31' and I got:

1399 blks/sec ext3-ordered !!!

It turns out that this kernel ignores the O_SYNC on open... ha.


P.S.  one other note.  Usually I let this my benchmark run for a few 
minutes, and yet, when the individual process scores (4 process test case) 
print out, I see that one process has made almost, or exactly 0 progress.  
0 blocks written in a minute.

| David Mansfield              |
| david cobite com             |

[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]