Messages in this thread | | | Date | Fri, 18 May 2007 09:59:43 +0200 | From | Nick Piggin <> | Subject | Re: [rfc] increase struct page size?! |
| |
On Fri, May 18, 2007 at 12:43:04AM -0700, Andrew Morton wrote: > On Fri, 18 May 2007 09:32:23 +0200 Nick Piggin <npiggin@suse.de> wrote: > > > On Fri, May 18, 2007 at 12:19:05AM -0700, Andrew Morton wrote: > > > On Fri, 18 May 2007 06:08:54 +0200 Nick Piggin <npiggin@suse.de> wrote: > > > > > > > Many batch operations on struct page are completely random, > > > > > > But they shouldn't be: we should aim to place physically contiguous pages > > > into logically contiguous pagecache slots, for all the reasons we > > > discussed. > > > > For big IO batch operations, pagecache would be more likely to be > > physically contiguous, as would LRU, I suppose. > > read(), write(), truncate(), writeback, pagefault. Pretty common stuff.
Of course, but if you're doing them on random-ish ranges, or multiple files, or continually on the same file while parts of it get reclaimed and reinstantiated...
> > I'm more thinking of operations where things get reclaimed over time, > > touched or dirtied in slightly different orderings, interleaved with > > other allocations, etc. > > Yes, that can happen. But in such cases we by definition aren't touching > the pageframes very often. I'd assert that when the kernel is really > hitting those pageframes hard, it is commonly doing this in ascending > pagecache order.
I'm not sure that I would always agree. Sure if there is random *IO* involved, then it is going to be slow and we won't be hitting page frames so hard. But if there is random pagecache access, it will hit them almost as hard.
> > > If/when that happens, there will be a *lot* of locality of reference > > > against the pageframes in a lot of important codepaths. > > > > And when it doesn't happen, we eat 75% more cache misses. And for that > > matter we eat 75% more cache misses for non-batch operations like > > allocating or freeing a page by slab, for example. > > "measure twice, cut once"
Definitely agree there.
- 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/
| |