Re: [PATCH -v2 -mm] add extra free kbytes tunable

[David Rientjes <rientjes@google.com>]

On Wed, 12 Oct 2011, Rik van Riel wrote:

> > > The problem is that we may be dealing with bursts, not steady
> > > states of allocations.  Without knowing the size of a burst,
> > > we have no idea when we should wake up kswapd to get enough
> > > memory freed ahead of the application's allocations.
> > 

Raising the priority of kswapd to be the highest possible when triggered 
by rt-tasks should help to reclaim memory faster.  If that doesn't work 
fully with Con's patch on Satoru's testcase then we'll want to extend it 
to raise the priority for a running kswapd when a higher priority thread 
calls into the page allocator slowpath.  If that also doesn't mitigate the 
problem entirely, then we'll need to suggest raising min_free_kbytes so 
these threads have a larger pool of exclusive access to memory when the 
burst first happens.

> > That problem remains with this patch - it just takes a larger burst.
> > 
> > Unless the admin somehow manages to configure the tunable large enough
> > to cover the largest burst, and there aren't other applications
> > allocating memory during that burst, and the time between bursts is
> > sufficient for kswapd to be able to sufficiently replenish free-page
> > reserves.  All of which sounds rather unlikely.
> 
> It depends on the system. For a setup which is packed to
> the brim with workloads, this patch is not likely to help.
> On the other hand, on a system that is packed to the brim
> with workloads, you are unlikely to get low latencies anyway.
> 
> For situations where people really care about low latencies,
> I imagine having dedicated hardware for a workload is not at
> all unusual, and the patch works for that.
> 

If it's dedicated hardware, then you should be able to just raise 
min_free_kbytes so that rt-tasks get exclusive access to a larger amount 
of memory.