Re: [RFC] Avoiding fragmentation through different allocator

[Mel Gorman]

On Wed, 12 Jan 2005, Matt Mackall wrote:

> On Wed, Jan 12, 2005 at 09:09:24PM +0000, Mel Gorman wrote:
> > I stress-tested this patch very heavily and it never oopsed so I am
> > confident of it's stability, so what is left is to look at the results of
> > this patch were and I think they look promising in a number of respects. I
> > have graphs that do not translate to text very well, so I'll just point you
> > to http://www.csn.ul.ie/~mel/projects/mbuddy-results-1 instead.
>
> This graph rather hard to comprehend.
>

There is not a lot of ways to illustrate how pages are allocated
throughout an address space. If you have a better suggestion, I can update
the relevant tool.

> > The results were not spectacular but still very interesting. Under heavy
> > stresing (updatedb + 4 simultaneous -j4 kernel compiles with avg load 15)
> > fragmentation is consistently lower than the standard allocator. It could
> > also be a lot better if there was some means of purging caches, userpages
> > and buffers but thats in the future. For the moment, the only real control
> > I had was the buffer pages.
>
> You might stress higher order page allocation with a) 8k stacks turned
> on b) UDP NFS with large read/write.
>

CONFIG_4KSTACKS is not set on my system so I'm already using 8k stacks.
Based on Trond's mail, I checked my network and the MTU is 1500 bytes so I
don't think I can force high-order allocations there.

Also, while I agree this test would be important, I think it'll be way
more important when there is something in place that actually tries to
free up high-order blocks and then compare.

-- 
Mel Gorman
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/