Hi Michal!
On 6/27/23 18:02, Neha Ojha wrote:
Hi Michal,
Thank you for volunteering to help test the Reef release!
On Tue, Jun 27, 2023 at 6:44 AM Michal Strnad
<michal.strnad@xxxxxxxxx> wrote:
Hi everyone,
We read that you are looking for ceph users who would be willing
to help
with performance testing of a new version of Ceph called Reef. We
would
like to volunteer and offer our assistance :-).
Currently, we are setting up a large cluster consisting of fifty
storage
nodes, each with 24 rotational disks and 8 NVMe drives, some of which
are designated for Bluestore and others for data purposes. Each of
these
machines is equipped with an AMD EPYC 7282 16-Core processor,
~314GB of
memory, and a 2x25Gbps network connection. The network on each of
these
machines is used for both public and cluster communication, and if
necessary, we can prioritize one over the other through QoS
adjustments
within the VLAN. However, we haven't had the need to do so thus far.
Furthermore, we have sixteen application servers for monitors, MGR,
metadata servers, and radosgw gateways. Each of these application
servers is equipped with an AMD EPYC 7502 32-Core processor,
~250GB of
memory, and a 2x25Gbps network connection.
Both the storage and application servers are connected to two
Nexus 9000
switches with connectivity reaching several 100Gbps towards the
internet.
The mentioned cluster will be operational within a few weeks, with
Ceph
already installed and ready to undergo performance testing. Once
this is
ready, it will be possible to start testing the Reef version. We
anticipate having approximately 2-3 weeks for testing. Are you
interested in the performance results? To achieve better results, it
would be beneficial to coordinate these tests in some way, so that we
don't repeat what others have already tried. Could you please
guide us
on what specific aspects we should focus on, which parameters to
test,
and how to properly conduct the tests?
We are particularly interested to see the performance impact of the
new RockDB version we'll be shipping with Reef. I am adding Mark to
this email to provide guidance on performance tests.
As Neha mentioned, we both did a fairly major rocksdb upgrade and also
changed Ceph's default bluestore RocksDB tunings after many years. We
*think* this is generally going to be an improvement, but there are some
trade offs. We expect lower write latency, higher write IOPS, and
reduced CPU consumption in the kv sync thread (this is a bigger deal on
all-flash setups), but there's a chance that we might also see higher
write-amplification on the DB device in some scenarios. In the testing
I did I saw lower write amp in some tests and higher write amp in
others. Paul Cuzner also ran tests over at IBM and saw higher CPU usage
during reads in some tests, though I think in the most recent version of
the changes those improved and he also saw the write latency
improvement. His tests were primarily focused on low-medium load
testing. Initially I was doing max load testing but after seeing his
results also ran some low-load tests.
I suspect that with HDD+Flash setups none of this is going to actually
matter that much as the rotational latency of the HDDs is likely to be
the dominating effect. It would be nice to verify that though. Given
that this is a new cluster that you guys are setting up, we won't have
any kind of historical data about existing RocksDB behavior to compare
with. We can still look at the general behavior of RocksDB though and
see if there's anything obviously wrong with it. I don't think we'll
necessarily need cluster access unless something really interesting pops
up, but perhaps Neha and Paul have something they want to look at.
As far as workloads go, I'd say if you guys have any real workloads you
intend to do (veaam, scientific apps, etc) that might actually be the
most worthwhile. We do a lot of synthetic testing internally. Real
workloads on real clusters are harder for us to do ourselves. Heavy
write workloads (especially smaller writes or smaller objects) would
probably be the most worthwhile for this round given the changes we're
interested in looking at.
Mark
After an agreement, it will be possible to arrange some form of
access
to the machines, for example, by meeting via video conference and
fine-tuning them together. Alternatively, we can also work on it
through
email, IRC, Slack, or any other suitable means.
We are coordinating community efforts around such testing in
#ceph-at-scale slack channel in ceph-storage.slack.com
<http://ceph-storage.slack.com>. I sent you an invite.
Thanks,
Neha
Kind regards,
Michal Strnad
On 6/13/23 22:27, Neha Ojha wrote:
> Hi everyone,
>
> This is the first release candidate for Reef.
>
> The Reef release comes with a new RockDB version (7.9.2) [0], which
> incorporates several performance improvements and features. Our
internal
> testing doesn't show any side effects from the new version, but
we are very
> eager to hear community feedback on it. This is the first
release to have
> the ability to tune RockDB settings per column family [1], which
allows for
> more granular tunings to be applied to different kinds of data
stored in
> RocksDB. A new set of settings has been used in Reef to optimize
> performance for most kinds of workloads with a slight penalty in
some
> cases, outweighed by large improvements in use cases such as
RGW, in terms
> of compactions and write amplification. We would highly
encourage community
> members to give these a try against their performance benchmarks
and use
> cases. The detailed list of changes in terms of RockDB and
BlueStore can be
> found in https://pad.ceph.com/p/reef-rc-relnotes.
>
> If any of our community members would like to help us with
performance
> investigations or regression testing of the Reef release
candidate, please
> feel free to provide feedback via email or in
> https://pad.ceph.com/p/reef_scale_testing. For more active
discussions,
> please use the #ceph-at-scale slack channel in
ceph-storage.slack.com <http://ceph-storage.slack.com>.
>
> Overall things are looking pretty good based on our testing.
Please try it
> out and report any issues you encounter. Happy testing!
>
> Thanks,
> Neha
>
> Get the release from
>
> * Git at git://github.com/ceph/ceph.git
<http://github.com/ceph/ceph.git>
> * Tarball at https://download.ceph.com/tarballs/ceph-18.1.0.tar.gz
> * Containers at https://quay.io/repository/ceph/ceph
> * For packages, see
https://docs.ceph.com/en/latest/install/get-packages/
> * Release git sha1: c2214eb5df9fa034cc571d81a32a5414d60f0405
>
> [0] https://github.com/ceph/ceph/pull/49006
> [1] https://github.com/ceph/ceph/pull/51821
> _______________________________________________
> ceph-users mailing list -- ceph-users@xxxxxxx
> To unsubscribe send an email to ceph-users-leave@xxxxxxx
--
Best Regards,
Mark Nelson
Head of R&D (USA)
Clyso GmbH
p: +49 89 21552391 12
a: Loristraße 8 | 80335 München | Germany
w: https://clyso.com | e: mark.nelson@xxxxxxxxx
We are hiring: https://www.clyso.com/jobs/
_______________________________________________
Dev mailing list -- dev@xxxxxxx
To unsubscribe send an email to dev-leave@xxxxxxx
