Hello Christian.
2017-08-24 22:43 GMT-03:00 Christian Balzer <chibi@xxxxxxx>:
Hello,
On Thu, 24 Aug 2017 14:49:24 -0300 Guilherme Steinmüller wrote:
> Hello Christian.
>
> First of all, thanks for your considerations, I really appreciate it.
>
> 2017-08-23 21:34 GMT-03:00 Christian Balzer <chibi@xxxxxxx>:
>
> >
> > Hello,
> >
> > On Wed, 23 Aug 2017 09:11:18 -0300 Guilherme Steinmüller wrote:
> >
> > > Hello!
> > >
> > > I recently installed INTEL SSD 400GB 750 SERIES PCIE 3.0 X4 in 3 of my
> > OSD
> > > nodes.
> > >
> > Well, you know what's coming now, don't you?
> >
> > That's a consumer device, with 70GB writes per day endurance.
> > unless you're essentially having a read-only cluster, you're throwing away
> > money.
> >
>
> Yes, we knew that we were going to buy a consumer device due to our limited
> budget and our objective of constructing a small plan of a production
> cloud. This model seemed acceptable. It was the top list of the consumer
> models on Sebastien's benchmarks
>
> We are a lab that depends on different budget sources to accquire
> equipments, so they can vary and most of the time we are limited by
> different budget ranges.
>
Noted, I hope your tests won't last too long or move a lot of data. ^o^
> Dell has a parameter called *Cache Line Size: 32 bytes* and HP doesn't have
> >
> > > First of all, here's is an schema describing how my cluster is:
> > >
> > > [image: Imagem inline 1]
> > >
> > > [image: Imagem inline 2]
> > >
> > > I primarily use my ceph as a beckend for OpenStack nova, glance, swift
> > and
> > > cinder. My crushmap is configured to have rulesets for SAS disks, SATA
> > > disks and another ruleset that resides in HPE nodes using SATA disks too.
> > >
> > > Before installing the new journal in HPE nodes, i was using one of the
> > > disks that today are OSDs (osd.35, osd.34 and osd.33). After upgrading
> > the
> > > journal, i noticed that a dd command writing 1gb blocks in openstack nova
> > > instances doubled the throughput but the value expected was actually 400%
> > > or 500% since in the Dell nodes that we have another nova pool the
> > > throughput is around this value.
> > >
> > Apples, oranges and bananas.
> > You're comparing different HW (and no, I'm not going to look this up)
> > which may or may not have vastly different capabilities (like HW cache),
> > RAM and (unlikely relevant) CPU.
> >
>
>
> Indeed, we took this into account. The HP server were cheaper and have a
> poor configuration due that limited budget source.
>
>
> > Your NVMe may also be plugged into a different, insufficient PCIe slot for
> > all we know.
> >
>
> I checked this. I compared the slots identifying the slot information
> between the 3 dell nodes and 3 hp nodes by running:
>
> # ls -l /sys/block/nvme0n1
> # lspci -vvv -s 0000:06:00.0 <- slot identifier
>
> The only difference is:
>
> this.
>
That shouldn't be relevant, AFAIK.
>
>
> > You're also using very different HDDs, which definitely will be a factor.
> >
> >
> I thought that the backend disks would not interfer that much. For example,
> the ceph journal has a parameter called filestore max sync interval, which
> means that ceph journal will commit the transactions to the backend OSDs in
> a defined interval, ours is set to 35. So the client requests go first to
> SSD and than is commited to the OSDs.
>
As I wrote before, the journal comes not into play for any large amounts
of data unless massively tuned and/or under extreme pressure.
You need to touch much more of the journal and filestore parameters than
max_sync, which will do nothing to prevent from min_sync and other values
to start flushing more or less immediately.
And tuning things so the journal is used extensively by default will
result in I/O storms slowing things to a crawl when it finally flushes to
the HDDs.
If your google foo is strong enough you should find the relevant
discussions about this, often in context with SSD OSDs where such tuning
makes some sense.
So the Dells do have a HW cache, which of course will help immensely.
>
> > But most importanly you're comparing 2 pools of vastly different ODS
> > count, no wonder a pool with 15 OSDs is faster in sequential writes than
> > one with 9.
> >
> > Here is a demonstration of the scenario and the difference in performance
> > > between Dell nodes and HPE nodes:
> > >
> > >
> > >
> > > Scenario:
> > >
> > >
> > > - Using pools to store instance disks for OpenStack
> > >
> > >
> > > - Pool nova in "ruleset SAS" placed on c4-osd201, c4-osd202 and
> > > c4-osd203 with 5 osds per hosts
> > >
> > SAS
> > >
> > > - Pool nova_hpedl180 in "ruleset NOVA_HPEDL180" placed on
> > c4-osd204,
> > > c4-osd205, c4-osd206 with 3 osds per hosts
> > >
> > SATA
> > >
> > > - Every OSD has one partition of 35GB in a INTEL SSD 400GB 750
> > > SERIES PCIE 3.0 X4
> > >
> > Overkill, but since your NVMe will die shortly anyway...
> >
> > With large sequential tests, the journal will have nearly NO impact on the
> > result, even if tuned to that effect.
> >
> > >
> > > - Internal link for cluster and public network of 10Gbps
> > >
> > >
> > > - Deployment via ceph-ansible. Same configuration define in
> > ansible
> > > for every host on cluster
> > >
> > >
> > >
> > > *Instance on pool nova in ruleset SAS:*
> > >
> > >
> > > # dd if=/dev/zero of=/mnt/bench bs=1G count=1 oflag=direct
> > > 1+0 records in
> > > 1+0 records out
> > > 1073741824 bytes (1.1 GB, 1.0 GiB) copied, 2.56255 s, 419 MB/s
> > >
> > This is a very small test for what you're trying to determine and not
> > going to be very representative.
> > If for example there _is_ a HW cache of 2GB on the Dell nodes, it would
> > fit nicely in there.
> >
> >
> Dell has PERC H730 Mini (Embedded) each with cache memory size of 1024 MB
> otherwise my HP uses a B140i dynamic array. Both HP and Dell doesn't use
> any raid level for the OSDs, just Dell for the Operating System.
>
Basically it boils down to what you're trying to test/compare here:
>
>
> > >
> > > *Instance on pool nova in ruleset NOVA_HPEDL180:*
> > >
> > > # dd if=/dev/zero of=/mnt/bench bs=1G count=1 oflag=direct
> > > 1+0 records in
> > > 1+0 records out
> > > 1073741824 bytes (1.1 GB, 1.0 GiB) copied, 11.8243 s, 90.8 MB/s
> > >
> > >
> > > I made some FIO benchmarks as suggested by Sebastien (
> > > https://www.sebastien-han.fr/blog/2014/10/10/ceph-how-to-
> > > test-if-your-ssd-is-suitable-as-a-journal-device/ ) and the command
> > with 1
> > > job returned me about 180MB/s of throughput in recently installed nodes
> > > (HPE nodes). I made some hdparm benchmark in all SSDs and everything
> > seems
> > > normal.
> > >
> > I'd consider a 180MB/s result from a device that supposedly does 900MB/s a
> > fail, but then again those tests above do NOT reflect journal usage
> > reality but a more of a hint if something is totally broken or not.
> >
> > >
> > > I can't see what is causing this difference of throughput since the
> > network
> > > is not a problem and i think that cpu and memory are not crucial since i
> > > was monitoring the cluster with atop command and i didn't notice
> > saturation
> > > of resources. My only though is that I have less workload in
> > nova_hpedl180
> > > pool in HPE nodes and less disks per node and this ca influence in the
> > > throughput of the journal.
> > >
> > How busy are your NVMe journals during that test on the Dells and HPs
> > respectively?
> > Same for the HDDs.
> >
>
>
> I can't say it now precisely, but what I can tell you for sure is that
> monitoring these two pools, both thoughtput and disk usage, I can see that
> the workload for the pool that is placed on the Dell nodes is significantly
> higher than the pool in the HP node. For example, the OSDs in the Dell node
> often keep the usage between 70% and 100%, different than HP OSDs, that
> vary between 10% and 40%.
>
1) The speed of your NVMe journal devices?
Put OSDs on them (with inline journal) and run extensive tests.
And with that I mean fio with 4MB for sequential speeds, 4K for IOPS and
latency, direct, sync, etc.
2) The actual production speed of these different servers and their pools?
Same as above and stated before, run _long_ tests and see where speeds
stabilize and what the utilizations are during that time.
And as always, sequential speed is in nearly all use cases not what
matters in Ceph cluster anyway, certainly not one serving VMs.
Certainly after this discussion I can see the big picture better. I will plan with my team some representative tests.
Cheers
Christian
>
> >
> > Again, run longer, larger tests to get something that will actually
> > register, also atop with shorter intervals.
> >
> > Christian
> > >
> > > Any clue about what is missing or what is happening?
> > >
> > > Thanks in advance.
> >
> >
> > --
> > Christian Balzer Network/Systems Engineer
> > chibi@xxxxxxx Rakuten Communications
> >
--
Christian Balzer Network/Systems Engineer
chibi@xxxxxxx Rakuten Communications
_______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
