lkml.org 
[lkml]   [2003]   [Mar]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] concurrent block allocation for ext2 against 2.5.64
William Lee Irwin III <wli@holomorphy.com> wrote:
>> dbench on 32x/48G NUMA-Q, aic7xxx adapter, pbay disk, 32K PAGE_SIZE
>> (pgcl was used for benchmark feasibility purposes)
>> throughput:
>> ----------
>> before:
>> Throughput 61.5376 MB/sec 512 procs
>> dbench 512 637.21s user 15739.41s system 565% cpu 48:16.28 total
>> after:
>> Throughput 104.074 MB/sec 512 procs
>> (GRR, didn't do time, took ca. 30 minutes)

On Fri, Mar 14, 2003 at 08:54:55PM -0800, Andrew Morton wrote:
> `dbench 512' will presumably do lots of IO and spend significant
> time in I/O wait. You should see the effects of this change more
> if you use fewer clients (say, 32) so it doesn't hit disk.
> On quad power4, dbench 32:

Hmm. I'm just trying to spawn enough tasks to keep the cpus busy to get
a large enough thread pool to have something to run when someone sleeps.
There's enough idle time now that this sounds like the wrong direction
to move the task count in...


On Fri, Mar 14, 2003 at 08:54:55PM -0800, Andrew Morton wrote:
> Unpatched:
> Throughput 334.372 MB/sec (NB=417.965 MB/sec 3343.72 MBit/sec)
[...]
> patched:
> Throughput 335.262 MB/sec (NB=419.078 MB/sec 3352.62 MBit/sec)
[...]
> No difference at all.
> On the quad Xeon (after increasing dirty_ratio and dirty_background_ratio so
> I/O was negligible) I was able to measure a 1.5% improvement.
> I worry about the hardware you're using there.

Why? The adapter is "vaguely modern" (actually acquired as part of a
hunt for an HBA w/a less buggy driver) but the box and disks and so on
are still pretty ancient, so the absolute numbers aren't useful.

To get a real comparison we'd have to compare spindles, HBA's, and
cpus, and attempt to factor them out. The disks are actually only
capable of doing 30MB/s or 40MB/s, the buses can only do 40MB/s, and
the cpus are 700MHz P-III's. Where dbench gets its numbers faster than
wirespeed I have no idea...

This locking issue may just need more cpus to bring out.


William Lee Irwin III <wli@holomorphy.com> wrote:
>> profile:
>> --------
[...]
>> after:
>> vma samples %-age symbol name
>> c0106ff4 52751908 30.8696 default_idle
>> c01dc3b0 28988721 16.9637 __copy_to_user_ll
>> c01dc418 8240854 4.82242 __copy_from_user_ll
>> c011e472 8044716 4.70764 .text.lock.fork
[...]
> Lots of idle time. Try it with a smaller client count, get the I/O out of
> the picture.

I'll have trouble as there won't be enough tasks to keep the cpus busy.
Why do you think reducing the client count gets io out of the picture?
Why do you think reducing the client count will reduce idle time?


William Lee Irwin III <wli@holomorphy.com> wrote:
>> vmstat (short excerpt, edited for readability):

On Fri, Mar 14, 2003 at 08:54:55PM -0800, Andrew Morton wrote:
> With what interval?

Sorry, 1s.


William Lee Irwin III <wli@holomorphy.com> wrote:
>> after:
>> procs -----------memory---------- -----io---- --system-- ----cpu----
>> r b free buff cache bi bo in cs us sy id wa
>> 60 11 38659840 533920 9226720 100 1672 2760 1853 5 66 11 18
>> 31 23 38565472 531264 9320384 240 1020 1195 1679 2 35 37 26
>> 23 23 38384928 521952 9503104 772 3372 5624 5093 2 62 9 27
>> 24 31 37945664 518080 9916448 1536 5808 10449 13484 1 45 13 41
>> 31 86 37755072 516096 10091104 1040 1916 3672 9744 2 51 15 32
>> 24 30 37644352 512864 10192960 900 1612 3184 8414 2 49 12 36
>> There's a lot of odd things going on in both of the vmstat logs.

On Fri, Mar 14, 2003 at 08:54:55PM -0800, Andrew Morton wrote:
> Where are all those interrupts coming from?

Well, the timer interrupt is a killer. 1KHz*num_cpus_online() blows
goats for sufficiently large num_cpus_online(), but for some reason
things are slower without it. I suspect that scheduling response
time is somehow dependent on it.

I got a hold of an aic7xxx so io throughput is slightly better than my
usual NUMA-Q runs (i.e. oopsen). The disks are still clockwork, though.


-- wli
-
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 13:33    [W:0.074 / U:2.436 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site