Re: Kernel 2.6.9 Multiple Page Allocation Failures

[Lukas Hejtmanek <xhejtman@hell.sks3.muni.cz>]

On Thu, Dec 09, 2004 at 09:29:29PM +1100, Nick Piggin wrote:
> Perhaps they weren't working properly in 2.6.6, or something else is 
> buggy in
> 2.6.9. 16MB in 2.6.9 should definitely give a a larger atomic reserve 
> than 900K
> in 2.6.6... so if it isn't an issue with network buffer behaviour, then the
> only other possibility AFAIKS is that page reclaim latency or efficiency has
> got much worse. This would be unlikely as it should cause problems and
> regressions on other workloads.

Increasing TCP buffers in 2.6.6 works somehow because it eliminates high RTT
over the large distance.

Here is exaclty what I do:
/sbin/sysctl -w net/core/rmem_max=8388608
/sbin/sysctl -w net/core/wmem_max=8388608
/sbin/sysctl -w net/core/rmem_default=1048576
/sbin/sysctl -w net/core/wmem_default=1048576
/sbin/sysctl -w net/ipv4/tcp_window_scaling=1
/sbin/sysctl -w net/ipv4/tcp_rmem="4096 1048576 8388608"
/sbin/sysctl -w net/ipv4/tcp_wmem="4096 1048576 8388608"
echo 32768 > /proc/sys/vm/min_free_kbytes
/sbin/ifconfig eth0 txqueuelen 1000

-- 
Lukáš Hejtmánek