[lkml]   [2002]   [Jul]   [22]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH][1/2] return values shrink_dcache_memory etc
"Martin J. Bligh" wrote:
> >> > If we can get something in place which works acceptably on Martin
> >> > Bligh's machines, and we can see that the gains of rmap (whatever
> >> > they are ;)) are worth the as-yet uncoded pains then let's move on.
> >> > But until then, adding new stuff to the VM just makes a `patch -R'
> >> > harder to do.
> >>
> >> I have the same kinds of machines and have already been testing with
> >> precisely the many tasks workloads he's concerned about for the sake of
> >> correctness, and efficiency is also a concern here. highpte_chain is
> >> already so high up on my priority queue that all other work is halted.
> >
> > OK. But we're adding non-trivial amounts of new code simply
> > to get the reverse mapping working as robustly as the virtual
> > scan. And we'll always have rmap's additional storage requirements.
> >
> > At some point we need to make a decision as to whether it's all
> > worth it. Right now we do not even have the information on the
> > pluses side to do this. That's worrisome.
> These large NUMA machines should actually be rmap's glory day in the
> sun.

"should be". Sigh. Be nice to see an "is" one day ;)

> Per-node kswapd, being able to free mem pressure on one node
> easily (without cross-node bouncing), breakup of the lru list into
> smaller chunks, etc. These actually fix some of the biggest problems
> that we have right now and are hard to solve in other ways.
> The large rmap overheads we still have to kill seem to me to be the
> memory usage and the fork overhead. There's also a certain amount of
> overhead to managing any more data structures, of course. I think we
> know how to kill most of it. I don't think adding highpte_chain is
> the correct thing to do ... seems like adding insult to injury. I'd
> rather see us drive a silver stake through the problem's heart and
> kill it properly ...

Well that would be nice. And by extension, pte-highmem gets a stake
as well.

Do you think that large pages alone would be enough to allow us
to leave pte_chains (and page tables?) in ZONE_NORMAL, or would
shared pagetables also be needed?

Was it purely Oracle which drove pte-highmem, or do you think
that page table and pte_chain consumption could be a problem
on applications which can't/won't use large pages?

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to
More majordomo info at
Please read the FAQ at

 \ /
  Last update: 2005-03-22 13:22    [W:0.078 / U:7.936 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site