Comments on: Linux Kernel Tuning

By: Russ

Russ — Sat, 14 Nov 2009 17:20:50 +0000

It’s worth noting that this setting was picked when we were running something fairly near the beginning of the 2.6 kernel tree (2.6.5-2.6.10). I believe the overcommit logic has been tweaked quite a lot since then, and it may no longer be correct.

At last.fm we only changed this on our netboot setup, which has no swap. I think in general, always allowing overcommit is a sensible move to reduce complexity on these systems. (We would find it more preferable to see an OOM situation on the machine than to see undefined userland behaviour due to failing mallocs.)

By: jau

jau — Thu, 12 Nov 2009 13:39:11 +0000

Why did you choose vm.overcommit = 1? Did the default heuristics handling fail?

By: Biff Martin

Biff Martin — Mon, 23 Mar 2009 21:17:35 +0000

Try these settings. You will have huge buffers.

net.ipv4.tcp_timestamps = 0
net.ipv4.tcp_sack = 0
net.ipv4.tcp_rmem = 210000 210000 210000
net.ipv4.tcp_wmem = 210000 210000 210000
net.ipv4.tcp_mem = 210000 210000 210000

net.core.rmem_max = 1073741824
net.core.wmem_max = 1073741824
net.core.rmem_default = 1073741824
net.core.wmem_default = 1073741824
net.core.optmem_max = 524287
net.core.netdev_max_backlog = 30000000

By: Biff Martin

Biff Martin — Mon, 23 Mar 2009 21:15:59 +0000

How about tuning the scheduler? do you have any advice?

Biff

By: imissmyjuno

imissmyjuno — Fri, 13 Mar 2009 05:02:14 +0000

I have a feeling this will come in handy someday. In the meantime, I must tell you the stylesheet of your blog is very confusing – the subheaders within every post are larger and are bold and the post titles are smaller and have no font weight. A bit hard to tell if one is reading the same or a new post.

Love Last.Fm :D