vm behaviour in 2.4 (was Re: disk performance still wrecked in 2.4!)

[Jens Axboe]

On Fri, Jun 30 2000, Michal Ostrowski wrote:
> 
> I mentioned a similar problem about a month ago.  Jens Axboe provided
> a patch which re-worked the elevator code and restored performance to
> 2.2 levels.  I think you'll be able to find his patches in the
> "people" directory.

That particular patch has been included in the stock kernel. While
dbench now gives much better results than 2.2, tests like tiobench
and bonnie[++] show that we're still not as good as 2.2 wrt I/O.

Try running vmstat 1 while doing disk benches (or just copying files),
and the vm looks to be behaving rather oddly:

[snipped]

axboe@bart:/test > dd if=test_file of=/dev/null bs=8192
axboe@bart:~ > vmstat 1
    memory    swap         io     system         cpu
buff  cache  si  so    bi    bo   in    cs  us  sy  id
216  56604   6   9   175   136  149   143   1   3  96
216  56604   0   0     0     0  103    10   0   0 100
132  56692 244 544  7821   151  655   731   2  56  42
128  56980 196 128  6435    37  554   596   0  47  53
124  57404   8  20   774     6  156    80   0   6  94
128  57044 176 120  3505    33  343   326   0  20  80
132  57156  80 140  5082    40  465   473   1  35  64
128  57716  88 196  8810    59  712   818   3  61  35
128  57136  96 240  8752    70  709   807   1  59  41
128  57768  16  88  1193    23  191   126   1  10  89
128  56832 148 304  8257    87  687   789   0  63  37
128  57116  68 284  8224    81  677   780   1  58  41
124  57216  96 336  8382    95  698   809   0  67  33

There's a significant amount of swap in/out

Same test without swap enabled:

axboe@bart:~ > vmstat 1
   memory    swap         io     system         cpu
buff  cache  si  so    bi    bo   in    cs  us  sy  id
212   2172   7  10   211   125  149   137   1   3  96
220  18240   0   0  4024     6  375   426   1  26  73
244  39104   0   0  5222     0  434   549   0  33  67
 88  53440   0   0  6018    24  501   602   0  65  35
 80  53244   0   0  7530    12  580   685   1  72  27
 84  53368   0   0  6905    10  541   613   1  72  27
 84  53632   0   0  5014     8  425   474   2  52  46
 84  53660   0   0  6153    14  505   610   0  66  34
 76  53772   0   0  2690     6  277   259   0  32  68
 76  53812   0   0  1826     4  223   176   0  23  77
 80  53284   0   0  5652    14  470   519   0  75  25
 84  53752   0   0  2583     3  270   252   0  29  71
 88  53336   0   0  5128     6  433   488   0  50  50

While (obviously) no swaps occur here, the bi rate is also
much slower. So I'm hoping vm guys still have a bit
of coding and tweaking to do ;)

For comparison, 2.2.14 shows _very_ consistent rates and with
very little occasional swapping.

-- 
* Jens Axboe <[e-mail]>
* SuSE Labs