On Monday 05 August 2002 01:33, Andrew Morton wrote:
> OK, I warmed the code up a bit and did some more measurement.
> Your locking patch has improved things significantly. And
> we're now getting hurt by all the cache misses walking the
> pte chains.
Well that's a relief. I was beginning to believe that your 4 way has some
sort of built-in anti-optimization circuit.
Are you still seeing no improvement on four way smp?
> ...And it's all really in __page_remove_rmap, kmem_cache_alloc/free.
>
> If we convert the pte_chain structure to
>
> struct pte_chain {
> struct pte_chain *next;
> pte_t *ptes[L1_CACHE_BYTES - 4];
> };
>
> and take care to keep them compacted we shall reduce the overhead
> of both __page_remove_rmap and the slab functions by up to 7, 15
> or 31-fold, depending on the L1 size. page_referenced() wins as well.
>
> Plus we almost halve the memory consumption of the pte_chains
> in the high sharing case. And if we have to kmap these suckers
> we reduce the frequency of that by 7x,15x,31x,etc.
>
> I'll code it tomorrow.
Sounds good. There is still some tuning to be done on the rmap lock
batching, to distribute the locks better and set anon page->indexes more
intelligently. I expect this to be good for another percent or two, nothing
really exciting.
-- Daniel - 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 : Wed Aug 07 2002 - 22:00:26 EST