On Tue, 22 Oct 2002, Dave McCracken wrote:
>
> --On Tuesday, October 22, 2002 12:02:27 -0700 "Martin J. Bligh"
> <mbligh@aracnet.com> wrote:
>
>
> >Bill Davidsen wrote:
> >> I'm just trying to decide what this might do for a news server with
> >> hundreds of readers mmap()ing a GB history file. Benchmarks show the 2.5
> >> has more latency the 2.4, and this is likely to make that more obvious.
> >
> > On the other hand, I don't think shared pagetables have an mmap hook,
> > though that'd be easy enough to add. And if you're not reading the whole
> > history file, presumably the PTEs will only be sparsely instantiated
> > anyway.
>
> Actually shared page tables work on any shared memory area, no matter how
> it was created. When a page fault occurs and there's no pte page already
> allocated (the common case for any newly mapped region) it checks the vma
> to see if it's shared. If it's shared, it gets the address_space for that
> vma, then walks through all the shared vmas looking for one that's mapped
> at the same address and offset and already has a pte page that can be
> shared.
That's more encouraging.
> So if your history file is mapped at the same address for all your
> processes then it will use shared page tables. While it might be a nice
> add-on to allow sharing if they're mapped on the same pte page boundary,
> that doesn't seem likely enough to justify the extra work.
The reader processes are either forked (INN w/ daemon nnrpd) or pthreads
(twister, earthquake, diablo). So everything will be the same.
Another thought, how does this play with NUMA systems? I don't have the
problem, but presumably there are implications.
-- bill davidsen <davidsen@tmr.com> CTO, TMR Associates, Inc Doing interesting things with little computers since 1979.- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Thu Oct 31 2002 - 22:00:23 EST