[PATCH 00/11] makedumpfile: Add zstd support for makedumpfile
HAGIO KAZUHITO(萩尾 一仁)
k-hagio-ab at nec.com
Wed Sep 22 01:16:11 PDT 2021
Hi Tao Liu,
Merged them into a patch and applied:
https://github.com/makedumpfile/makedumpfile/commit/afd0a6db2a0543217f8e46955a1b44b71f7e7ef3
Thanks,
Kazu
> -----Original Message-----
> Hi Tao Liu,
>
> -----Original Message-----
> > Hello Kazu,
> >
> > Sorry for the late reply.
> >
> > On Fri, Sep 17, 2021 at 07:03:50AM +0000, HAGIO KAZUHITO(萩尾 一仁) wrote:
> > > -----Original Message-----
> > > > -----Original Message-----
> > > > > > > > This patch set adds ZSTD compression support to makedumpfile. With ZSTD compression
> > > > > > > > support, the vmcore dump size and time consumption can have a better balance than
> > > > > > > > zlib/lzo/snappy.
> > > > > > > >
> > > > > > > > How to build:
> > > > > > > >
> > > > > > > > Build using make:
> > > > > > > > $ make USEZSTD=on
> > > > > > > >
> > > > > > > > Performance Comparison:
> > > > > > > >
> > > > > > > > How to measure
> > > > > > > >
> > > > > > > > I took a x86_64 machine which had 4T memory, and the compression level
> > > > > > > > range from (-3 to 4) for ZSTD, as well as zlib/lzo/snappy compression.
> > > > > > > > All testing was done by makedumpfile single thread mode.
> > > > > > > >
> > > > > > > > As for compression performance testing, in order to avoid the performance
> > > > > > > > bottle neck of disk I/O, I used the following makedumpfile cmd, which took
> > > > > > > > lzo compression as an example. "--dry-run" will not write any data to disk,
> > > > > > > > "--show-stat" will output the vmcore size after compression, and the time
> > > > > > > > consumption can be collected from the output logs.
> > > > > > > >
> > > > > > > > $ makedumpfile -d 0 -l /proc/kcore vmcore --dry-run --show-stat
> > > > > > > >
> > > > > > > >
> > > > > > > > As for decompression performance testing, I only tested the (-d 31) case,
> > > > > > > > because the vmcore size of (-d 0) case is too big to fit in the disk, in
> > > > > > > > addtion, to read a oversized file from disk will encounter the disk I/O
> > > > > > > > bottle neck.
> > > > > > > >
> > > > > > > > I triggered a kernel crash and collected a vmcore. Then I converted the
> > > > > > > > vmcore into specific compression format using the following makedumpfile
> > > > > > > > cmd, which would get a lzo format vmcore as an example:
> > > > > > > >
> > > > > > > > $ makedumpfile -l vmcore vmcore.lzo
> > > > > > > >
> > > > > > > > After all vmcores were ready, I used the following cmd to perform the
> > > > > > > > decompression, the time consumption can be collected from the logs.
> > > > > > > >
> > > > > > > > $ makedumpfile -F vmcore.lzo --dry-run --show-stat
> > > > > > > >
> > > > > > > >
> > > > > > > > Result charts
> > > > > > > >
> > > > > > > > For compression:
> > > > > > > >
> > > > > > > > makedumpfile -d31 | makedumpfile -d0
> > > > > > > > Compression time vmcore size | Compression time vmcore size
> > > > > > > > zstd-3 325.516446 5285179595 | 8205.452248 51715430204
> > > > > > > > zstd-2 332.069432 5319726604 | 8057.381371 51732062793
> > > > > > > > zstd-1 309.942773 5730516274 | 8138.060786 52136191571
> > > > > > > > zstd0 439.773076 4673859661 | 8873.059963 50993669657
> > > > > > > > zstd1 406.68036 4700959521 | 8259.417132 51036900055
> > > > > > > > zstd2 397.195643 4699263608 | 8230.308291 51030410942
> > > > > > > > zstd3 436.491632 4673306398 | 8803.970103 51043393637
> > > > > > > > zstd4 543.363928 4668419304 | 8991.240244 51058088514
> > > > > > > > zlib 561.217381 8514803195 | 14381.755611 78199283893
> > > > > > > > lzo 248.175953 16696411879 | 6057.528781
> 90020895741
> > > > > > > > snappy 231.868312 11782236674 | 5290.919894 245661288355
> > > > > > > >
> > > > > > > > For decompression:
> > > > > > > >
> > > > > > > > makedumpfile -d31
> > > > > > > > decompress time vmcore size
> > > > > > > > zstd-3 477.543396 5289373448
> > > > > > > > zstd-2 478.034534 5327454123
> > > > > > > > zstd-1 459.066807 5748037931
> > > > > > > > zstd0 561.687525 4680009013
> > > > > > > > zstd1 547.248917 4706358547
> > > > > > > > zstd2 544.219758 4704780719
> > > > > > > > zstd3 555.726343 4680009013
> > > > > > > > zstd4 558.031721 4675545933
> > > > > > > > zlib 630.965426 8555376229
> > > > > > > > lzo 427.292107 16849457649
> > > > > > > > snappy 446.542806 11841407957
> > > > > > > >
> > > > > > > > Discussion
> > > > > > > >
> > > > > > > > For zstd range from -3 to 4, compression level 2 (ZSTD_dfast) has
> > > > > > > > the best time consumption and vmcore dump size balance.
> > > > > > >
> > > > > > > Do you have a result of -d 1 compression test? I think -d 0 is not
> > > > > > > practical, I would like to see a -d 1 result of such a large vmcore.
> > > > > > >
> > > > > >
> > > > > > No, I haven't tested the -d 1 case. I have returned the machine which used
> > > > > > for performance testing, I will borrow and test on it again, please wait for
> > > > > > a while...
> > > > >
> > > > > Thanks, it would be helpful.
> > > > >
> > > > > >
> > > > > > > And just out of curiosity, what version of zstd are you using?
> > > > > > > When I tested zstd last time, compression level 1 was faster than 2, iirc.
> > > > > > >
> > > > > >
> > > > > > The OS running on the machine is fedora34, I used its default zstd package, whose
> > > > > > version is v1.4.9.
> > > > >
> > > > > Thanks for the info.
> > > > >
> > > > > >
> > > > > > > btw, ZSTD_dfast is an enum of ZSTD_strategy, not for compression level?
> > > > > >
> > > > > > Yes, it's enum of ZSTD_strategy [1].
> > > > >
> > > > > ok, so it'll have to be replaced with "2" to avoid confusion.
> > > > >
> > > > > >
> > > > > > [1]: https://zstd.docsforge.com/dev/api-documentation/#advanced-compression-api-requires-v140
> > > > > >
> > > > > > Thanks,
> > > > > > Tao Liu
> > > > > >
> > > > > > > (no need to update for now, I will review later)
> > > > >
> > > > > The series almost looks good to me (though I will merge those into a patch),
> > > > > just two questions are:
> > > > > - whether 2 is the best balanced compression level,
> > >
> > > As far as I've tested on two machines this time, compression level 1 was faster
> > > than 2. There is no large difference between them, but generally 1 should be
> > > faster than 2 according to the zstd manual:
> > > "The lower the level, the faster the speed (at the cost of compression)."
> > > And as you know, level 0 is unstable, that was the same when I tested before.
> > >
> > > So currently I would prefer 1 rather than 2, what do you think?
> >
> > As we mentioned before, I have tested the -d 1 compression measurement on
> > the same x86_64 machine with 4T memory:
> >
> > compression time| vmcore size
> > zstd-3 4620.795194 31720632985
> > zstd-2 4545.636437 31716847503
> > zstd-1 4516.076298 32113300399
> > zstd0 4663.17618 30967496299
> > zstd1 4618.386313 31010305809
> > zstd2 4633.535771 31005073344
> > zstd3 4673.240663 30967855841
> > zstd4 4771.1416 30965914853
> > lzo 4801.958368 34920417584
> > zlib 4442.257105 43482765168
> > snappy 4433.957005 38594790371
> >
> > As for the decompression, I didn't get a meaningful value, because the vmcore size
> > were too large, and the most time was spent on disk I/O, thus the decompression time
> > measurement didn't show an obvious difference.
> >
> > I agree the compression level 1 and 2 don't have a big difference. I'm OK with
> > your preference.
>
> OK, thank you for the testing! As we have tested, the compression level 1
> and 2 often change places each other, let's choose 1 for general speed.
>
> I will merge the series with some adjustments, please wait for a while.
>
> >
> > >
> > > Results:
> > > * RHEL8.4 with libzstd-1.4.4 / 64GB filled with QEMU memory/images mainly
> > > # free
> > > total used free shared buff/cache available
> > > Mem: 65599824 21768028 549364 4668 43282432 43078828
> > > Swap: 32964604 4827916 28136688
> > >
> > > makedumpfile -d 1 makedumpfile -d 31
> > > copy sec. write bytes copy sec. write bytes
> > > zstd1 220.979689 26456659213 9.014176 558845000
> > > zstd2 227.774602 26402437190 9.078599 560681256
> > > lzo 83.406291 33078995065 3.603778 810219860
> > >
> > > * RHEL with libzstd-1.5.0 / 64GB filled with kernel source code mainly
> > > # free
> > > total used free shared buff/cache available
> > > Mem: 65329632 9925536 456020 53086068 54948076 1549088
> > > Swap: 32866300 1607424 31258876
> > >
> > > makedumpfile -d 1 makedumpfile -d 31
> > > zstd1 520.844189 15537080819 53.494782 1200754023
> > > zstd2 533.912451 15469575651 53.641510 1199561609
> > > lzo 233.370800 20780821165 23.281374 1740041245
> > >
> > > (Used /proc/kcore, so not stable memory, but measured zstd 3 times and
> > > picked the middle elapsed time.)
> > >
> >
> > Thanks for sharing the results. Just for curiosity, can you share your way
> > of testing as well? I can improve mine for later use.
>
> I did these:
> (1) fill memory
> - the former case: on a KVM host running over 100 days with 8 guests
> $ cat guest.img > /dev/null for some images until swap is used.
> - the latter case: tar xf linux-5.14.tar.xz and copy multiple times
> into /dev/shm until swap is used.
>
> (2) run makedumpfile for /proc/kcore
> I also used the command you used and picked up the elapsed time of
> "Copying data", to measure only the compression time, and size of
> "Write bytes".
>
> # makedumpfile -z -d 1 --dry-run --show-stat /proc/kcore vmcore
> ...
> ...[Copying data ] : 9.463352 seconds
> ...
> Write bytes : 535151008
>
> Thanks,
> Kazu
>
> _______________________________________________
> kexec mailing list
> kexec at lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/kexec
More information about the kexec
mailing list