Re: [PATCH v7 0/5] Update LZ4 compressor module

[Sven Schmidt]

On Mon, Feb 13, 2017 at 09:03:24AM +0900, Minchan Kim wrote:
> Hi Sven,
> 
> On Sun, Feb 12, 2017 at 12:16:17PM +0100, Sven Schmidt wrote:
> > 
> > 
> > 
> > On 02/10/2017 01:13 AM, Minchan Kim wrote:
> > > Hello Sven,
> > >
> > > On Thu, Feb 09, 2017 at 11:56:17AM +0100, Sven Schmidt wrote:
> > >> Hey Minchan,
> > >>
> > >> On Thu, Feb 09, 2017 at 08:31:21AM +0900, Minchan Kim wrote:
> > >>> Hello Sven,
> > >>>
> > >>> On Sun, Feb 05, 2017 at 08:09:03PM +0100, Sven Schmidt wrote:
> > >>>>
> > >>>> This patchset is for updating the LZ4 compression module to a version based
> > >>>> on LZ4 v1.7.3 allowing to use the fast compression algorithm aka LZ4 fast
> > >>>> which provides an "acceleration" parameter as a tradeoff between
> > >>>> high compression ratio and high compression speed.
> > >>>>
> > >>>> We want to use LZ4 fast in order to support compression in lustre
> > >>>> and (mostly, based on that) investigate data reduction techniques in behalf of
> > >>>> storage systems.
> > >>>>
> > >>>> Also, it will be useful for other users of LZ4 compression, as with LZ4 fast
> > >>>> it is possible to enable applications to use fast and/or high compression
> > >>>> depending on the usecase.
> > >>>> For instance, ZRAM is offering a LZ4 backend and could benefit from an updated
> > >>>> LZ4 in the kernel.
> > >>>>
> > >>>> LZ4 homepage: http://www.lz4.org/
> > >>>> LZ4 source repository: https://github.com/lz4/lz4
> > >>>> Source version: 1.7.3
> > >>>>
> > >>>> Benchmark (taken from [1], Core i5-4300U @1.9GHz):
> > >>>> ----------------|--------------|----------------|----------
> > >>>> Compressor      | Compression  | Decompression  | Ratio
> > >>>> ----------------|--------------|----------------|----------
> > >>>> memcpy          |  4200 MB/s   |  4200 MB/s     | 1.000
> > >>>> LZ4 fast 50     |  1080 MB/s   |  2650 MB/s     | 1.375
> > >>>> LZ4 fast 17     |   680 MB/s   |  2220 MB/s     | 1.607
> > >>>> LZ4 fast 5      |   475 MB/s   |  1920 MB/s     | 1.886
> > >>>> LZ4 default     |   385 MB/s   |  1850 MB/s     | 2.101
> > >>>>
> > >>>> [1] http://fastcompression.blogspot.de/2015/04/sampling-or-faster-lz4.html
> > >>>>
> > >>>> [PATCH 1/5] lib: Update LZ4 compressor module
> > >>>> [PATCH 2/5] lib/decompress_unlz4: Change module to work with new LZ4 module version
> > >>>> [PATCH 3/5] crypto: Change LZ4 modules to work with new LZ4 module version
> > >>>> [PATCH 4/5] fs/pstore: fs/squashfs: Change usage of LZ4 to work with new LZ4 version
> > >>>> [PATCH 5/5] lib/lz4: Remove back-compat wrappers
> > >>>
> > >>> Today, I did zram-lz4 performance test with fio in current mmotm and
> > >>> found it makes regression about 20%.
> > >>>
> > >>> "lz4-update" means current mmots(git://git.cmpxchg.org/linux-mmots.git) so
> > >>> applied your 5 patches. (But now sure current mmots has recent uptodate
> > >>> patches)
> > >>> "revert" means I reverted your 5 patches in current mmots.
> > >>>
> > >>>                      revert    lz4-update
> > >>>
> > >>>       seq-write       1547       1339      86.55%
> > >>>      rand-write      22775      19381      85.10%
> > >>>        seq-read       7035       5589      79.45%
> > >>>       rand-read      78556      68479      87.17%
> > >>>    mixed-seq(R)       1305       1066      81.69%
> > >>>    mixed-seq(W)       1205        984      81.66%
> > >>>   mixed-rand(R)      17421      14993      86.06%
> > >>>   mixed-rand(W)      17391      14968      86.07%
> > >>
> > >> which parts of the output (as well as units) are these values exactly?
> > >> I did not work with fio until now, so I think I might ask before misinterpreting my results.
> > >
> > > It is IOPS.
> > >
> > >>  
> > >>> My fio description file
> > >>>
> > >>> [global]
> > >>> bs=4k
> > >>> ioengine=sync
> > >>> size=100m
> > >>> numjobs=1
> > >>> group_reporting
> > >>> buffer_compress_percentage=30
> > >>> scramble_buffers=0
> > >>> filename=/dev/zram0
> > >>> loops=10
> > >>> fsync_on_close=1
> > >>>
> > >>> [seq-write]
> > >>> bs=64k
> > >>> rw=write
> > >>> stonewall
> > >>>
> > >>> [rand-write]
> > >>> rw=randwrite
> > >>> stonewall
> > >>>
> > >>> [seq-read]
> > >>> bs=64k
> > >>> rw=read
> > >>> stonewall
> > >>>
> > >>> [rand-read]
> > >>> rw=randread
> > >>> stonewall
> > >>>
> > >>> [mixed-seq]
> > >>> bs=64k
> > >>> rw=rw
> > >>> stonewall
> > >>>
> > >>> [mixed-rand]
> > >>> rw=randrw
> > >>> stonewall
> > >>>
> > >>
> > >> Great, this makes it easy for me to reproduce your test.
> > >
> > > If you have trouble to reproduce, feel free to ask me. I'm happy to test it. :)
> > >
> > > Thanks!
> > >
> > 
> > Hi Minchan,
> > 
> > I will send an updated patch as a reply to this E-Mail. Would be really grateful If you'd test it and provide feedback!
> > The patch should be applied to the current mmots tree.
> > 
> > In fact, the updated LZ4 _is_ slower than the current one in kernel. But I was not able to reproduce such large regressions
> > as you did. I now tried to define FORCE_INLINE as Eric suggested. I also inlined some functions which weren't in upstream LZ4,
> > but are defined as macros in the current kernel LZ4. The approach to replace LZ4_ARCH64 with the function call _seemed_ to behave
> > worse than the macro, so I withdrew the change.
> > 
> > The main difference is, that I replaced the read32/read16/write... etc. functions using memcpy with the other ones defined 
> > in upstream LZ4 (which can be switched using a macro). 
> > The comment of the author stated, that they're as fast as the memcpy variants (or faster), but not as portable
> > (which does not matter since we're not dependent for multiple compilers).
> > 
> > In my tests, this version is mostly as fast as the current kernel LZ4.
> 
> With a patch you sent, I cannot see enhancement so I wanted to dig in and
> found how I was really careless.
> 
> I have tested both test with CONFIG_KASAN. OMG. With disabling it, I don't
> see any regression any more. So, I'm really really *sorry* about noise and
> wasting your time. However, I am curious why KASAN makes such difference.
> 

Hey Minchan,

I'm glad to hear that! Nevertheless, the changes discussed here made some differences in my own tests (I believe it got a bit
faster now) and we have the functions properly inlined, where this makes sense. Also, I added the '-O3' C-flag as Eric suggested.
So, this was not really a waste of time, I think.

> The reason I tested new updated lz4 is description says lz4 fast and
> want to use it in zram. How can I do that? and How faster it is compared
> to old?
>

Unfortunately, in the current implementation (in crypto/lz4.c, which is used by zram) I'm setting the acceleration parameter 
(which is the paramer making the compression 'fast', see LZ4_compress_fast) to 1 (which is the default) since I did not know how this
patchset is accepted and this equals the behaviour currently available in kernel.

Basically, the logic is 'higher acceleration = faster compression = lower compression ratio' and vice versa. 
I included some benchmarks in my patch 0/5 E-Mail taken from the official LZ4:

> > >>>> ----------------|--------------|----------------|----------
> > >>>> Compressor      | Compression  | Decompression  | Ratio
> > >>>> ----------------|--------------|----------------|----------
> > >>>> memcpy          |  4200 MB/s   |  4200 MB/s     | 1.000
> > >>>> LZ4 fast 50     |  1080 MB/s   |  2650 MB/s     | 1.375
> > >>>> LZ4 fast 17     |   680 MB/s   |  2220 MB/s     | 1.607
> > >>>> LZ4 fast 5      |   475 MB/s   |  1920 MB/s     | 1.886
> > >>>> LZ4 default     |   385 MB/s   |  1850 MB/s     | 2.101
> > >>>>

fast 50 means: acceleration=50, default: acceleration=1.

Besides the proposed patchset, I tried to implement a module parameter in crypto/lz4.c to set the acceleration factor.
In my tests, the module parameter works out great.
But I think this is subject to a future, separate patch. Especially since I had to 'work around' the crypto/testmgr.c, 
which only tests acceleration=1 and there's no limit for acceleration.  

Thanks for your help,

Sven