Re: [PATCH] memcpy_flushcache: use cache flusing for larger lengths

From: Dan Williams
Date: Wed Apr 08 2020 - 15:29:34 EST


On Wed, Apr 8, 2020 at 11:54 AM Mikulas Patocka <mpatocka@xxxxxxxxxx> wrote:
>
>
>
> On Tue, 7 Apr 2020, Dan Williams wrote:
>
> > On Tue, Apr 7, 2020 at 8:02 AM Mikulas Patocka <mpatocka@xxxxxxxxxx> wrote:
> > >
> > > [ resending this to x86 maintainers ]
> > >
> > > Hi
> > >
> > > I tested performance of various methods how to write to optane-based
> > > persistent memory, and found out that non-temporal stores achieve
> > > throughput 1.3 GB/s. 8 cached stores immediatelly followed by clflushopt
> > > or clwb achieve throughput 1.6 GB/s.
> > >
> > > memcpy_flushcache uses non-temporal stores, I modified it to use cached
> > > stores + clflushopt and it improved performance of the dm-writecache
> > > target significantly:
> > >
> > > dm-writecache throughput:
> > > (dd if=/dev/zero of=/dev/mapper/wc bs=64k oflag=direct)
> > > writecache block size 512 1024 2048 4096
> > > movnti 496 MB/s 642 MB/s 725 MB/s 744 MB/s
> > > clflushopt 373 MB/s 688 MB/s 1.1 GB/s 1.2 GB/s
> > >
> > > For block size 512, movnti works better, for larger block sizes,
> > > clflushopt is better.
> >
> > This should use clwb instead of clflushopt, the clwb macri
> > automatically converts back to clflushopt if clwb is not supported.
>
> But we want to invalidate cache, we do not expect CPU to access these data
> anymore (it will be accessed by a DMA engine during writeback).

The cluflushopt and clwb instructions should have identical overhead,
but clwb wins on the rare chance the written data is needed again
soon. If it is never needed again then the cost of dropping a clean
cache line is the same as if the line was invalidated in the first
instance. In both cases (clflushopt and clwb) the snoop traffic
overhead is still paid whether the written-back line is still present
in the cache or not.

>
> > > I was also testing the novafs filesystem, it is not upstream, but it
> > > benefitted from similar change in __memcpy_flushcache and
> > > __copy_user_nocache:
> > > write throughput on big files - movnti: 662 MB/s, clwb: 1323 MB/s
> > > write throughput on small files - movnti: 621 MB/s, clwb: 1013 MB/s
> > >
> > >
> > > I submit this patch for __memcpy_flushcache that improves dm-writecache
> > > performance.
> > >
> > > Other ideas - should we introduce memcpy_to_pmem instead of modifying
> > > memcpy_flushcache and move this logic there? Or should I modify the
> > > dm-writecache target directly to use clflushopt with no change to the
> > > architecture-specific code?
> >
> > This also needs to mention your analysis that showed that this can
> > have negative cache pollution effects [1], so I'm not sure how to
> > decide when to make the tradeoff. Once we have movdir64b the tradeoff
> > equation changes yet again:
> >
> > [1]: https://lore.kernel.org/linux-nvdimm/alpine.LRH.2.02.2004010941310.23210@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/
>
> I analyzed it some more. I have created this program that tests writecache
> w.r.t. cache pollution:
>
> http://people.redhat.com/~mpatocka/testcases/pmem/misc/l1-test-2.c
>
> It fills the cache with a chain of random pointers and then walks these
> pointers to evaluate cache pollution. Between the walks, it writes data to
> the dm-writecache target.
>
> With the original kernel, the result is:
> 8503 - 11366
> real 0m7.985s
> user 0m0.585s
> sys 0m7.390s
>
> With dm-writecache hacked to use cached writes + clflushopt:
> 8513 - 11379
> real 0m5.045s
> user 0m0.670s
> sys 0m4.365s
>
> So, the hacked dm-writecache is significantly faster, while the cache
> micro-benchmark doesn't show any more cache pollution.

Nice. These are now the pmem numbers, or dram? Otherwise, what changed
that was making nt-writes on pmem perform better compared to your
previous test? I'm just trying to track the results.

> That's for dm-writecache. Are there some other significant users of
> memcpy_flushcache that need to be checked?

The only other user is direct and dax-I/O to the pmem driver.