On 06/26/17 11:33, Denis V. Lunev wrote:
On 06/26/2017 12:20 PM, Peter Lieven wrote:
Am 26.06.2017 um 10:28 schrieb Kevin Wolf:
[ Cc: qemu-devel; don't post to qemu-block only! ]
Am 26.06.2017 um 09:57 hat Peter Lieven geschrieben:
Hi,
I am currently working on optimizing speed for compressed QCOW2
images. We use them for templates and would also like to use them for
backups, but the latter is almost infeasible because using gzip for
compression is horribly slow. I tried to experiment with different
options to deflate, but in the end I think its better to use a
different compression algorithm for cases where speed matters. As we
already have probing for it in configure and as it is widely used I
would like to use LZO for that purpose. I think it would be best to
have a flag to indicate that compressed blocks use LZO compression,
but I would need a little explaination which of the feature fields I
have to use to prevent an older (incompatible) Qemu opening LZO
compressed QCOW2 images.
I also have already some numbers. I converted a fresh Debian 9 Install
which has an uncomressed QCOW2 size of 1158 MB with qemu-img to a
compressed QCOW2. With GZIP compression the result is 356MB whereas
the LZO version is 452MB. However, the current GZIP variant uses 35
seconds for this operation where LZO only needs 4 seconds. I think is
is a good trade in especially when its optional so the user can
choose.
What are your thoughts?
We had a related RFC patch by Den earlier this year, which never
received many comment and never got out of RFC:
https://lists.gnu.org/archive/html/qemu-devel/2017-03/msg04682.html
I was not aware of that one. Thanks for pointing out.
So he chose a different algorithm (zstd). When I asked, he posted a
comparison of algorithms (however a generic one and not measured in the
context of qemu) that suggests that LZO would be slightly faster, but
have a considerable worse compression ratio with the settings that were
benchmarked.
My idea to choose LZO was that it is widely available and available in
any distro you can think of. We already have probing for it in configure.
My concern with ZSTD would be that it seems there are no packages
available for most distros and that it seems to be multi-threaded. I
don't
know if this will cause any trouble?
We have had that compression working in multithreaded process.
I think it's clear that if there is any serious interest in compression,
we'll want to support at least one more algorithm. What we still need to
evaluate is which one(s) to take, and whether a simple incompatible flag
in the header like in Den's patch is enough or whether we should add a
whole new header field for the compression algorithm (like we already
have for encryption).
From my side there clearly is interest in optimizing the compression. Its
even possible to speed up zlib by 3-4x times by choosing other parameters
for deflate which unfortunately are not compatible with our inflate
settings.
I don't know if its worth creating a new header field. Even if we
spent to bits
in the end (one for LZO and one for ZSDT). I think this wouldn't hurt.
However,
there are likely to pop up new compression algorithms in the future and
a header would be more flexible.
I just don't want to make it too complicated and as you pointed out
compression is
not that interesting for most people - maybe due to its speed.
I think we need something generic but simple. I think that we should not
support compression with the different algorithm in the single file.
Speaking about compression, we do have different constraints for
different situation, f.e. backups are written once and rarely read while
generic compression in backing store is read frequently but never
read. Thus the exact algorithm should be selectable.
Pluggable / selectable compression methods are likely the most flexible
and future-proof. A new header sounds good to me (... said by someone
who comments on this from the sidelines.)
I would advise caution against multi-threaded compression libraries.
Unless they are coded very-very carefully with regard to signal handling
and general error handling / propagation, they cannot be considered
"opaque" enough.
(I had written and maintained the original (0.x) branch of "lbzip2",
which was extremely conscious of error handling and signals. That was a
challenge even in a standalone program, and I didn't even attempt to
retrofit the code to the existing libbz2 APIs (i.e. I never even tried
to librarize the code).)
This does not mean that people cannot get such a library right. It's
just that *by default* such a library will have a number of obscure bugs
related to: signals, forking, and general error handling. It could also
have problems with unbounded memory allocation. An MT compression
library that gets all of this right is the exception IMO, not the norm.
(I don't know anything about the ZSTD library; it could be such a high
quality library.)
Another complication with MT *de*compression is that the CPU demand from
the IO thread (which is by default responsible for handling IO, when not
using dataplane -- is that right?) would "leak" to other physical
processors. I believe this can interfere with use cases where people
carefully isolate host CPUs between "QEMU" and "non-QEMU" workloads,
plus pin QEMU's VCPU threads, and IO threads, to different host CPUs
(see vcpupin / emulatorpin / iothreadpin under
<http://libvirt.org/formatdomain.html#elementsCPUTuning>.) It's probably
possible to figure out the right thing for "ZSTD threads" as well, but
IMO it remains a complication nonetheless.
Personally I would recommend a new header, and LZO, as a starting point.
As pointed out above, LZO is widely available in distros. It has good
performance, and it is single-threaded similarly to zlib. I use LZO for
two QEMU-related purposes ATM:
- I use the kdump-lzo format when dumping guest memory
(virsh dump $DOMAIN $CORE_FLE --memory-only --format kdump-lzo)
- I use LZO compression for "virsh managedsave"
(by setting "save_image_format" in "/etc/libvirt/qemu.conf" to "lzop")