Andrew Tridgell wrote:
> Andrew,
> 
>  > So what you should do before generating the leak tool output is to put
>  > heavy memory pressure on the machine to try to get it to free up as much of
>  > that pagecache as possible.  bzero(malloc(lots)) will do it - create a real
>  > swapstorm, then do swapoff to kill remaining swapcache as well.
> 
> As you saw when you logged into the machine earlier tonight, when you
> suspend the dbench processes and run a memory filler the memory is
> reclaimed.
> 
> I still think its a bug though, as the oom killer is being triggered
> when it shouldn't be. I have 4G of ram in this machine, and I'm only
> running a couple of hundred processes that should be using maybe 500M
> in total, so for the oom killer to kick in might mean that the memory
> isn't being reclaimed under normal memory pressure. Certainly a ps
> shows no process using more than a few MB.
> 
> The oom killer report is below. This is with 2.6.11-rc2, with the pipe
> leak fix, and the pgown monitoring patch. It was running one nbench of
> size 50 and one dbench of size 40 at the time.
> 

There are various OOM killer improvements and fixes that have gone
into Andrew's kernel tree which should be included for 2.6.11.

I don't think the OOM killer was ever perfect in 2.6, but recent
tinkering in mm/ probably aggrivated it. *blush*

Here is another small OOM killer improvement. Previously we needed
to reclaim SWAP_CLUSTER_MAX pages in a single pass. That should be
changed so that we need only reclaim that many pages during the
entire try_to_free_pages run, without going OOM.

Andrea? Andrew? Look OK?