Re: 2.4.19rc2aa1 VM too aggressive?

[Andrea Arcangeli <andrea@suse.de>]

On Tue, Jul 23, 2002 at 01:22:35PM -0700, Stephen Hemminger wrote:
> When running sequential write performance tests on 2.4.19rc3ac3 I see a
> similar problem.  What happens is the page cache gets really big, then
> the machine starts swapping and becomes unusable.
> 
> It seems like there needs to be some upper bound on the page cache or
> flow control on file writes to not allow the cpu to get ahead of the
> disk. 

that's the write throttling.

> 
> This is vmstat output when running iozone, and it does first  .5 G file
> then a 1G file. The machine has lots of memory but when it fills, it
> goes off the deep end...
> 
>   procs                      memory    swap          io     system         cpu
>  r  b  w   swpd   free   buff  cache  si  so    bi    bo   in    cs  us  sy  id
>  1  0  1      0 234456  12128 542360   0   0     0 32475 1044    29   0  11  89
>  1  0  1      0 233744  12160 542360   0   0     0 62712 1065    79   0  18  82
>  0  1  0      0 303008  12160 542360   0   0     0 21751 1042    21   0   3  97
>  1  0  1      0 510140  12384 201112   0   0     0 11070 1042    18   0  14  86
>  0  1  0      0 302280  12660 542360   0   0     0 77826 1034    54   0  18  82
>  0  1  0      0 302280  12660 542360   0   0     0     0 1040     8   0   1  99
>  1  0  1      0 176144  12660 542360   0   0     0 39073 1043    23   0  12  88
>  0  1  0      0 302272  12660 542360   0   0     0 81925 1050    60   0  20  80
>  0  1  0      0 302268  12660 542360   0   0     0     0 1044     6   0   0 100
>  1  0  1      0 480496  12852 176408   0   0     0 10661 1043    18   0  14  86
>  0  1  0      0 301456  13156 542360   0   0     0 81511 1035    37   0  19  81
>  0  1  0      0 301452  13156 542360   0   0     0     0 1040     7   0   0 100
>  1  0  1      0 181724  13156 542360   0   0     0 30479 1047    18   0  10  90
>  0  1  0      0 301408  11864 543684   0   0     1 87331 1077    61   0  23  77
>  0  1  0      0 301400  11872 543684   0   0     0     2 1043     8   0   0 100
>  0  0  1      0 529624  11880 313604   0   0     0 26639 1043  1843   0   9  91
>  0  0  2      0 137892  11880 700356   0   0     0 92110 1036  2424   0  10  90
>  0  0  1      0  54688  11888 780868   0   0     0 29978 1038   515   0   2  98
>  0  1  1   7900   7024    976 848600 747 1926  1277 22411 1242   730   0   8  92
>  0  1  1   7900   7036    596 849352 1412 305  2351   738 1421   447   0   4  96
>  0  2  1   7900   7040    596 849488 1453 162  3077   174 1799   821   0  14  86
>  0  2  1   7900   7032    596 849772 888 150  1634   168 4111  1165   0  39  61

some seldom swapout is ok, the strange thing are those small
swapins/swapouts. I also assume it's writing using write(2), not with
map_shared+msync.

can you try:

	echo 1000 >/proc/sys/vm/vm_mapped_ratio

I also wonder if you've quite some amount of mapped address space durign
the benchmark. In such case there's no trivial way around it, the vm
will constantly found tons of mapped address space, and it will trigger
some swapouts, however the swapins shouldn't happen so fast in such
case.

In any case the sysctl will allow you to tune for your workload.

Andrea