On Sun, Aug 11, 2002 at 09:58:22PM -0700, Andrew Morton wrote:
> ext3 performs its own writeback alongside the core kernel's writeback
> decisions, so that complicates things.
I ran the test after mounting the partition as ext2 and saw a slight
decrease in performance (7-10 seconds over the duration of the test), but I
did not have time to run more than once so this could be a fluke. In general,
the `vmstat 1` output looked the same to me.
> > Results (average of 4 runs):
> >
> > 2.5.31-akpm: 2m 43s
> > 2.5.31: 2m 33s
> > 2.4.19: 2m 18s
>
> yes. For this workload (10 mbyte/sec ftp transfer onto a >20 meg/sec
> disk) the application should never block on IO - all writeback should
> happen via pdflush.
> You can make 2.5 use the 2.4 settings with
>
> cd /proc/sys/vm
> echo 30 > dirty_background_ratio
> echo 60 > dirty_async_ratio
> echo 70 > dirty_sync_ratio
These settings bring -akpm in line with stock 2.5.31, but they are both
still slower than 2.4.19 (which itself could do better, I think).
> and I expect you'll find that fixes it up. Setting dirty_background_ratio
> to 10% will make it even better. But it will hurt dbench numbers at
No real change at 10%. It's consistently a second or two faster than -akpm is at
30%, but not a drastic change.
> certain client counts, which is a national emergency.
>
> Sigh. I don't know what the right numbers are. There aren't any; that's
> the problem with magic numbers. That part of the kernel is making writeback
> and throttling decisions in total ignorance of the overall state of
> the system.
It certainly seems something is amiss. If we could actually manage to keep
the disk busy (and this is a fairly slow disk), we'd do wonderfully. But with
a 2-3 second pause every 4-5 seconds, we're transferring data barely 50% of the
time. (Yes, the pause is long enough the disk activity LED actually goes out.)
The short-term average transfer rate over the FTP connection is very
respectable for older hardware: 7-8 MB/s. But with the stalls, the overall
throughput is just over 4 MB/s.
> In fact, I'd be inclined to set the background ratio much lower than
> 2.4, and to hell with dbench. Because the lower level is better for
> real programs, as you've observed.
> Care to tune and retest?
Absolutely. I'll try whatever ideas/patches you want to throw at me.
BTW, full `vmstat 1` logs are available for all these tests if you want them.
--Adam
-
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/
This archive was generated by hypermail 2b29 : Thu Aug 15 2002 - 22:00:30 EST