Re: improving checksum cpu consumption in ksm

[Izik Eidus <ieidus@redhat.com>]

Hugh Dickins wrote:
> Yes, that's nice, thank you for looking into it.
>
> But please do some more along these lines, if you've time?
> Presumably the improvement from Jenkins lookup2 to lookup3
> is therefore more than 15%, but we cannot tell how much.
>
> I think you need to do a run with a null version of jhash2(),
> one just returning 0 or 0xffffffff (the first would settle down
> a little quicker because oldchecksum 0 will match the first time;
> but there should be no difference once you cut out settling time).
>
> And a run with an almost-null version of jhash2(), one which does
> also read the whole page sequentially into cache, so we can see
> how much is the processing and how much is the memory access.
>
> And also, while you're about it, a run with cmp_and_merge_page()
> stubbed out, so we can see how much is just the page table walking
> (and deduce from that how much is the radix tree walking and memcmping).
>
> Hmm, and a run to see how much is radix tree walking,
> by stubbing out the memcmping.
>
> Sorry... if you (or someone else following) have the time!
>   

Good ideas, The tests that I did were quick tests, I hope in Saturday 
(or maybe few days after it) I will have more time to spend on this area
I will try to collect and see how much every part is taking there.
So i will keep benchmarking it in terms of "how many loops it accomplish"

>
> Doesn't matter to your results, so long as it didn't crash;
> but I think you meant to say
>
>      p = (unsigned char *)(((unsigned long)p + 4095) & ~4095);
>      p_end = p + 1024 * 1024 * 100;
>   

Yes...


Thanks.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>