-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA256
Just swapping out spindles for SSD will not give you orders of magnitude performance gains as it does in regular cases. This is because Ceph has a lot of overhead for each I/O which limits the performance of the SSDs. In my testing, two Intel S3500 SSDs with an 8 core Atom (Intel(R) Atom(TM) CPU C2750 @ 2.40GHz) and size=1 and fio with 8 jobs and QD=8 sync,direct 4K read/writes produced 2,600 IOPs. Don't get me wrong, it will help, but don't expect spectacular results.
- ----------------
Robert LeBlanc
PGP Fingerprint 79A2 9CA4 6CC4 45DD A904 C70E E654 3BB2 FA62 B9F1
On Tue, Sep 1, 2015 at 8:01 AM, Kenneth Van Alstyne wrote:
Thanks for the awesome advice folks. Until I can go larger scale (50+ SATA disks), I’m thinking my best option here is to just swap out these 1TB SATA disks with 1TB SSDs. Am I oversimplifying the short term solution?
Thanks,
- --
Kenneth Van Alstyne
Systems Architect
Knight Point Systems, LLC
Service-Disabled Veteran-Owned Business
1775 Wiehle Avenue Suite 101 | Reston, VA 20190
c: 228-547-8045 f: 571-266-3106
www.knightpoint.com
DHS EAGLE II Prime Contractor: FC1 SDVOSB Track
GSA Schedule 70 SDVOSB: GS-35F-0646S
GSA MOBIS Schedule: GS-10F-0404Y
ISO 20000 / ISO 27001
Notice: This e-mail message, including any attachments, is for the sole use of the intended recipient(s) and may contain confidential and privileged information. Any unauthorized review, copy, use, disclosure, or distribution is STRICTLY prohibited. If you are not the intended recipient, please contact the sender by reply e-mail and destroy all copies of the original message.
On Aug 31, 2015, at 7:29 PM, Christian Balzer wrote:
Hello,
On Mon, 31 Aug 2015 12:28:15 -0500 Kenneth Van Alstyne wrote:
In addition to the spot on comments by Warren and Quentin, verify this by
watching your nodes with atop, iostat, etc.
The culprit (HDDs) should be plainly visible.
More inline:
Christian, et al:
Sorry for the lack of information. I wasn’t sure what of our hardware
specifications or Ceph configuration was useful information at this
point. Thanks for the feedback — any feedback, is appreciated at this
point, as I’ve been beating my head against a wall trying to figure out
what’s going on. (If anything. Maybe the spindle count is indeed our
upper limit or our SSDs really suck? :-) )
Your SSDs aren't the problem.
To directly address your questions, see answers below:
- CBT is the Ceph Benchmarking Tool. Since my question was more
generic rather than with CBT itself, it was probably more useful to post
in the ceph-users list rather than cbt.
- 8 Cores are from 2x quad core Intel(R) Xeon(R) CPU E5-2609 0 @
2.40GHz
Not your problem either.
- The SSDs are indeed Intel S3500s. I agree — not ideal, but
supposedly capable of up to 75,000 random 4KB reads/writes. Throughput
and longevity is quite low for an SSD, rated at about 400MB/s reads and
100MB/s writes, though. When we added these as journals in front of the
SATA spindles, both VM performance and rados benchmark numbers were
relatively unchanged.
The only thing relevant in regards to journal SSDs is the sequential write
speed (SYNC), they don't seek and normally don't get read either.
This is why a 200GB DC S3700 is a better journal SSD than the 200GB S3710
which is faster in any other aspect but sequential writes. ^o^
Latency should have gone down with the SSD journals in place, but that's
their main function/benefit.
- Regarding throughput vs iops, indeed — the throughput that I’m
seeing is nearly worst case scenario, with all I/O being 4KB block
size. With RBD cache enabled and the writeback option set in the VM
configuration, I was hoping more coalescing would occur, increasing the
I/O block size.
That can only help with non-SYNC writes, so your MySQL VMs and certain
file system ops will have to bypass that and that hurts.
As an aside, the orchestration layer on top of KVM is OpenNebula if
that’s of any interest.
It is actually, as I've been eying OpenNebula (alas no Debian Jessie
packages). However not relevant to your problem indeed.
VM information:
- Number = 15
- Worload = Mixed (I know, I know — that’s as vague of an answer
as they come) A handful of VMs are running some MySQL databases and
some web applications in Apache Tomcat. One is running a syslog
server. Everything else is mostly static web page serving for a low
number of users.
As others have mentioned, would you expect this load to work well with
just 2 HDDs and via NFS to introduce network latency?
I can duplicate the blocked request issue pretty consistently, just by
running something simple like a “yum -y update” in one VM. While that
is running, ceph -w and ceph -s show the following: root@dashboard:~#
ceph -s cluster f79d8c2a-3c14-49be-942d-83fc5f193a25 health HEALTH_WARN
1 requests are blocked > 32 sec
monmap e3: 3 mons at
{storage-1=10.0.0.1:6789/0,storage-2=10.0.0.2:6789/0,storage-3=10.0.0.3:6789/0}
election epoch 136, quorum 0,1,2 storage-1,storage-2,storage-3 osdmap
e75590: 6 osds: 6 up, 6 in pgmap v3495103: 224 pgs, 1 pools, 826 GB
data, 225 kobjects 2700 GB used, 2870 GB / 5571 GB avail
224 active+clean
client io 3292 B/s rd, 2623 kB/s wr, 81 op/s
[snip]
466 kB/s rd, 1863 kB/s wr, 148 op/s
This is a good sample, unless your reads can be satisfied from page cache
on your storage nodes or inside your VMs (more memory for the VMs may help
here), they are competing (seeks) with your write requests. So yeah, this
is probably as good as it gets.
I never seem to get anywhere near 300 op/s. If spindle count is indeed
the problem, is there anything else I can do to improve caching or I/O
coalescing to deal with my crippling IOP limit due to the low number of
spindles?
Other than replacing spindles with SSDs, not really.
Your client workload is too mixed for anything else but that or massively
more spindles.
On the other hand, I have a cluster with very few OSDs (4!), hundreds of
VMs and typical activities like this: 11750 kB/s wr, 1426 op/s.
Note the lack of writes, all these VMs run the same OS/application and are
basically write only.
Adding to that the OSDs are actually RAIDs behind a 4GB controller cache
and thus the "disks" aren't busy at all.
However reads, like rebooting VMs, impact this cluster quite a bit.
Christian
Thanks,
- --
Kenneth Van Alstyne
Systems Architect
Knight Point Systems, LLC
Service-Disabled Veteran-Owned Business
1775 Wiehle Avenue Suite 101 | Reston, VA 20190
c: 228-547-8045 f: 571-266-3106
www.knightpoint.com
DHS EAGLE II Prime Contractor: FC1 SDVOSB Track
GSA Schedule 70 SDVOSB: GS-35F-0646S
GSA MOBIS Schedule: GS-10F-0404Y
ISO 20000 / ISO 27001
Notice: This e-mail message, including any attachments, is for the sole
use of the intended recipient(s) and may contain confidential and
privileged information. Any unauthorized review, copy, use, disclosure,
or distribution is STRICTLY prohibited. If you are not the intended
recipient, please contact the sender by reply e-mail and destroy all
copies of the original message.
On Aug 31, 2015, at 11:01 AM, Christian Balzer wrote:
Hello,
On Mon, 31 Aug 2015 08:31:57 -0500 Kenneth Van Alstyne wrote:
Sorry about the repost from the cbt list, but it was suggested I post
here as well:
I wasn't even aware a CBT (what the heck does that acronym stand for?)
existed...
I am attempting to track down some performance issues in a Ceph
cluster recently deployed. Our configuration is as follows: 3
storage nodes,
3 nodes is, of course, bare minimum.
each with:
- 8 Cores
Of what, apples? Detailed information makes for better replies.
- 64GB of RAM
Ample.
- 2x 1TB 7200 RPM Spindle
Even if your cores where to be rotten apple ones, that's very few
spindles, so your CPU is unlikely to be the bottleneck.
- 1x 120GB Intel SSD
Details, again. From your P.S. I conclude that these are S3500's,
definitely not my choice for journals when it comes to speed and
endurance.
- 2x 10GBit NICs (In LACP Port-channel)
Massively overspec'ed considering your storage sinks/wells aka HDDs.
The OSD pool min_size is set to “1” and “size” is set to “3”. When
creating a new pool and running RADOS benchmarks, performance isn’t
bad — about what I would expect from this hardware configuration:
Rados bench uses by default 4MB "blocks", which is the optimum size for
(default) RBD pools.
Bandwidth does not equal IOPS (which are commonly measured in 4KB
blocks).
WRITES:
Total writes made: 207
Write size: 4194304
Bandwidth (MB/sec): 80.017
Stddev Bandwidth: 34.9212
Max bandwidth (MB/sec): 120
Min bandwidth (MB/sec): 0
Average Latency: 0.797667
Stddev Latency: 0.313188
Max latency: 1.72237
Min latency: 0.253286
RAND READS:
Total time run: 10.127990
Total reads made: 1263
Read size: 4194304
Bandwidth (MB/sec): 498.816
Average Latency: 0.127821
Max latency: 0.464181
Min latency: 0.0220425
This all looks fine, until we try to use the cluster for its purpose,
which is to house images for qemu-kvm, which are access using librbd.
Not that it probably matters, but knowing if this Openstack, Ganeti or
something else might be of interest.
I/O inside VMs have excessive I/O wait times (in the hundreds of ms at
times, making some operating systems, like Windows unusable) and
throughput struggles to exceed 10MB/s (or less). Looking at ceph
health, we see very low op/s numbers as well as throughput and the
requests blocked number seems very high. Any ideas as to what to look
at here?
Again, details.
How many VMs?
What are they doing?
Keep in mind that the BEST sustained result you could hope for here
(ignoring Ceph overhead and network latency) is the IOPS of 2 HDDs, so
about 300 IOPS at best. TOTAL.
health HEALTH_WARN
8 requests are blocked > 32 sec
monmap e3: 3 mons at
{storage-1=10.0.0.1:6789/0,storage-2=10.0.0.2:6789/0,storage-3=10.0.0.3:6789/0}
election epoch 128, quorum 0,1,2 storage-1,storage-2,storage-3 osdmap
e69615: 6 osds: 6 up, 6 in pgmap v3148541: 224 pgs, 1 pools, 819 GB
256 or 512 PGs would have been the "correct" number here, but that's of
little importance.
data, 227 kobjects 2726 GB used, 2844 GB / 5571 GB avail
224 active+clean
client io 3957 B/s rd, 3494 kB/s wr, 30 op/s
That's a lot of data being written for a tiny cluster like yours.
Looking at your nodes with atop or similar tools will likely reveal
that your HDDs are quite the busy beavers and can't keep up.
Also prolonged values from "ceph -w" might be educational.
Regards,
Christian
Of note, on the other list, I was asked to provide the following:
- ceph version 0.94.1
(e4bfad3a3c51054df7e537a724c8d0bf9be972ff)
- The SSD is split into 8GB partitions. These 8GB partitions
are used as journal devices, specified in /etc/ceph/ceph.conf. For
example: [osd.0] host = storage-1
osd journal
= /dev/mapper/INTEL_SSDSC2BB120G4_CVWL4363006R120LGNp1
- rbd_cache is enabled and qemu cache is set to “writeback"
- rbd_concurrent_management_ops is unset, so it appears the
default is “10”
Thanks,
- --
Kenneth Van Alstyne
Systems Architect
Knight Point Systems, LLC
Service-Disabled Veteran-Owned Business
1775 Wiehle Avenue Suite 101 | Reston, VA 20190
c: 228-547-8045 f: 571-266-3106
www.knightpoint.com
DHS EAGLE II Prime Contractor: FC1 SDVOSB Track
GSA Schedule 70 SDVOSB: GS-35F-0646S
GSA MOBIS Schedule: GS-10F-0404Y
ISO 20000 / ISO 27001
Notice: This e-mail message, including any attachments, is for the
sole use of the intended recipient(s) and may contain confidential and
privileged information. Any unauthorized review, copy, use,
disclosure, or distribution is STRICTLY prohibited. If you are not
the intended recipient, please contact the sender by reply e-mail and
destroy all copies of the original message.
- --
Christian Balzer Network/Systems Engineer
chibi@xxxxxxx Global OnLine Japan/Fusion Communications
http://www.gol.com/
- --
Christian Balzer Network/Systems Engineer
chibi@xxxxxxx Global OnLine Japan/Fusion Communications
http://www.gol.com/
_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
-----BEGIN PGP SIGNATURE-----
Version: Mailvelope v1.0.2
Comment: https://www.mailvelope.com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=52Nr
-----END PGP SIGNATURE-----
