On Tue, Nov 12, 2013 at 09:33:35AM -0800, Joe Perches wrote:
> On Tue, 2013-11-12 at 12:12 -0500, Neil Horman wrote:
> > On Mon, Nov 11, 2013 at 05:42:22PM -0800, Joe Perches wrote:
> > > Hi again Neil.
> > >
> > > Forwarding on to netdev with a concern as to how often
> > > do_csum is used via csum_partial for very short headers
> > > and what impact any prefetch would have there.
> > >
> > > Also, what changed in your test environment?
> > >
> > > Why are the new values 5+% higher cycles/byte than the
> > > previous values?
> > >
> > > And here is the new table reformatted:
> > >
> > > len set iterations Readahead cachelines vs cycles/byte
> > > 1 2 3 4 6 10 20
> > > 1500B 64MB 1000000 1.4342 1.4300 1.4350 1.4350 1.4396 1.4315
> > > 1.4555
> > > 1500B 128MB 1000000 1.4312 1.4346 1.4271 1.4284 1.4376 1.4318
> > > 1.4431
> > > 1500B 256MB 1000000 1.4309 1.4254 1.4316 1.4308 1.4418 1.4304
> > > 1.4367
> > > 1500B 512MB 1000000 1.4534 1.4516 1.4523 1.4563 1.4554 1.4644
> > > 1.4590
> > > 9000B 64MB 1000000 0.8921 0.8924 0.8932 0.8949 0.8952 0.8939
> > > 0.8985
> > > 9000B 128MB 1000000 0.8841 0.8856 0.8845 0.8854 0.8861 0.8879
> > > 0.8861
> > > 9000B 256MB 1000000 0.8806 0.8821 0.8813 0.8833 0.8814 0.8827
> > > 0.8895
> > > 9000B 512MB 1000000 0.8838 0.8852 0.8841 0.8865 0.8846 0.8901
> > > 0.8865
> > > 64KB 64MB 1000000 0.8132 0.8136 0.8132 0.8150 0.8147 0.8149
> > > 0.8147
> > > 64KB 128MB 1000000 0.8013 0.8014 0.8013 0.8020 0.8041 0.8015
> > > 0.8033
> > > 64KB 256MB 1000000 0.7956 0.7959 0.7956 0.7976 0.7981 0.7967
> > > 0.7973
> > > 64KB 512MB 1000000 0.7934 0.7932 0.7937 0.7951 0.7954 0.7943
> > > 0.7948
> > >
> >
> >
> > There we go, thats better:
> > len set iterations Readahead cachelines vs cycles/byte
> > 1 2 3 4 5 10 20
> > 1500B 64MB 1000000 1.3638 1.3288 1.3464 1.3505 1.3586 1.3527 1.3408
> > 1500B 128MB 1000000 1.3394 1.3357 1.3625 1.3456 1.3536 1.3400 1.3410
> > 1500B 256MB 1000000 1.3773 1.3362 1.3419 1.3548 1.3543 1.3442 1.4163
> > 1500B 512MB 1000000 1.3442 1.3390 1.3434 1.3505 1.3767 1.3513 1.3820
> > 9000B 64MB 1000000 0.8505 0.8492 0.8521 0.8593 0.8566 0.8577 0.8547
> > 9000B 128MB 1000000 0.8507 0.8507 0.8523 0.8627 0.8593 0.8670 0.8570
> > 9000B 256MB 1000000 0.8516 0.8515 0.8568 0.8546 0.8549 0.8609 0.8596
> > 9000B 512MB 1000000 0.8517 0.8526 0.8552 0.8675 0.8547 0.8526 0.8621
> > 64KB 64MB 1000000 0.7679 0.7689 0.7688 0.7716 0.7714 0.7722 0.7716
> > 64KB 128MB 1000000 0.7683 0.7687 0.7710 0.7690 0.7717 0.7694 0.7703
> > 64KB 256MB 1000000 0.7680 0.7703 0.7688 0.7689 0.7726 0.7717 0.7713
> > 64KB 512MB 1000000 0.7692 0.7690 0.7701 0.7705 0.7698 0.7693 0.7735
> >
> >
> > So, the numbers are correct now that I returned my hardware to its previous
> > interrupt affinity state, but the trend seems to be the same (namely that
> > there
> > isn't a clear one). We seem to find peak performance around a readahead of
> > 2
> > cachelines, but its very small (about 3%), and its inconsistent (larger set
> > sizes fall to either side of that stride). So I don't see it as a clear
> > win. I
> > still think we should probably scrap the readahead for now, just take the
> > perf
> > bits, and revisit this when we can use the vector instructions or the
> > independent carry chain instructions to improve this more consistently.
> >
> > Thoughts
>
> Perhaps a single prefetch, not of the first addr but of
> the addr after PREFETCH_STRIDE would work best but only
> if length is > PREFETCH_STRIDE.
>
> I'd try:
>
> if (len > PREFETCH_STRIDE)
> prefetch(buf + PREFETCH_STRIDE);
> while (count64) {
> etc...
> }
>
> I still don't know how much that impacts very short lengths.
>
> Can you please add a 20 byte length to your tests?
>
>
Sure, I modified the code so that we only prefetched 2 cache lines ahead, but
only if the overall length of the input buffer is more than 2 cache lines.
Below are the results (all counts are the average of 1000000 iterations of the
csum operation, as previous tests were, I just omitted that column).
len set cycles/byte cycles/byte improvement
no prefetch prefetch
===========================================================
20B 64MB 45.014989 44.402432 1.3%
20B 128MB 44.900317 46.146447 -2.7%
20B 256MB 45.303223 48.193623 -6.3%
20B 512MB 45.615301 44.486872 2.2%
1500B 64MB 1.364365 1.332285 1.9%
1500B 128MB 1.373945 1.335907 1.4%
1500B 256MB 1.356971 1.339084 1.2%
1500B 512MB 1.351091 1.341431 0.7%
9000B 64MB 0.850966 0.851077 -0.1%
9000B 128MB 0.851013 0.850366 0.1%
9000B 256MB 0.854212 0.851033 0.3%
9000B 512MB 0.857346 0.851744 0.7%
64KB 64MB 0.768224 0.768450 ~0%
64KB 128MB 0.768706 0.768884 ~0%
64KB 256MB 0.768459 0.768445 ~0%
64KB 512MB 0.768808 0.769404 -0.1%
The 20 byte results seem to have a few outliers. I'm guessing the improvement
came from good fortune in that the random selection happened to hit on the same
range of numbers a few times over, so we hit already cached data. I would
expect them to run more slowly (as the 2 and 3 rows illustrate), since 20B is
less than the 128 bytes in 2 cachelines on my test system, and so all were doing
is adding in an additional comparison and jump per iteration. Our sweet spot is
the 1500 byte range, giving us a small performance boost, but that quickly gets
lost in the noise as the buffer size grows beyond that.
I'm still left thinking we should just abandon the prefetch at this point and
keep the perf code until we have new instructions to help us with this further,
unless you see something I dont.
Neil
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
