Nick Piggin wrote:
> 
> 
> Mike Fedyk wrote:
> 
>> Nick Piggin wrote:
>>
>>>
>>>
>>> That is a much better sounding ratio. Of course that doesn't mean much
>>> if performance is worse. Slab might be getting reclaimed a little bit
>>> too hard vs pagecache now.
>>>
>>
>> I'll let you know.  My graphs are looking better, except for one 
>> instance of Xvnc (for one user -- I'm still tracking that one down) 
>> hitting a memory grabbing loop that made me kill it.
>>
> 
> Try to get /proc/meminfo and a sysrq + T trace if something like
> this happens.

I'll do that next time.

When did /proc/sysrq-trigger get in the kernel?  forgot about that at 
the time :(

> 
>>>> See:
>>>> http://www.matchmail.com/stats/lrrd/matchmail.com/srv-lnx2600.matchmail.com-memory.html 
>>>>
>>>>
>>>> Is there any way I can get the VM patches against 2.6.3?  I'm not 
>>>> comfortable with running -mm3 on this production server, especially 
>>>> seeing the "sync hang" bug.
>>>>
>>>
>>> Well your server wasn't going too badly with 2.6.3, wasn't it? Might
>>> as well just wait for them to get into the the tree.
>>
>>
>>
>> I might as well take out the third 512MB DIMM in that machine then...
>>
>> Any chance you could post a VM patch roll-up against 2.6.3 for little 
>> ole me?
>>
> 
> It is a bit easier said than done as you might have seen :P And
> I'm laz^W^W I happen to not agree with one of Andrew's patches,
> so it would go against all my principles ;)
> 
> IMO, shrink_slab-for-all-zones.patch and zone-balancing-fix.patch
> should be all you need although they won't shrink the slab as
> much as mm3. They should be pretty easy to port by hand.

OK, I'll give that a try.

Is the attached patch the latest version of your alternative patch 
instead of shrink_slab-for-all-zones.patch?