On Thu, 30 Apr 2015 18:01:44 -0400 Anthony Levesque wrote:
Im planning to setup 4-6 POC in the next 2 week to test various
scenarios here.
Im checking to get POC with s3610, s3710, p3500(seem to be knew. I know
the lifespam is lower) and maybe P3700
Don't ignore the S3700, it is faster in sequential writes than the 3710
because it uses older, less dense flash modules, thus more parallelism.
And with Ceph, especially when looking at the journals, you will hit the
max sequential write speed limit of the SSD long, loooong before you'll
hit the IOPS limit.
Both due to the nature of journal write and the little detail that you'll
hit the CPU performance wall before that.
The speed of the 400GB p3500 seem very nice and the price is alright.
The major difference will be the durability between the P3700 and P3500
and the IOPS.
Read the link below about write amplification, but that is something that
happens mostly on the OSD part, which in your case of 1TB EVOs is already
a scary prospect in my book.
in both option, they are the model with the lowest price per MB/s when
compare to the S series.
Price per MB/s is a good start, don't forget to factor in TBW/$ and try to
estimate what write loads your cluster will see.
But all of this is irrelevant if your typical write patterns will exceed
your CPU resources while your SSDs are bored.
For example this fio in a VM here:
---
# fio --size=4G --ioengine=libaio --invalidate=1 --direct=0 --numjobs=1 --rw=randwrite --name=fiojob --blocksize=4K --iodepth=32
write: io=1381.4MB, bw=16364KB/s, iops=4091 , runt= 86419msec
---
Will utilize all 8 3.1 GHz cores here, on a 3 node firefly cluster with 8
HDD OSDs and 4 journal SSDs (100GB S3700) per node.
While the journal SSDs are at 11% and the OSD HDDs at 30-40% utilization.
When changing that fio to direct=1, the IOPS drop to half of that.
With a block size of 4MB things of course change to the OSDs being 100%
busy, the SSDs about 60% (they can only do 200MB/s) and with 3-4 cores
worth being idle or in IOwait.
Model
Price per MB/s
DC S3500
120GB
$1.10
240GB
$1.01
300GB
$1.03
480GB
$1.28
DC S3610
200GB
$0.99
400GB
$1.14
480GB
$1.24
DC S3710
200GB
$1.17
DC P3500
400GB
$0.64
DC P3700
400GB
$0.96
As a side note, the expense doesn’t scare me directly, Its more that
we are going blind here since it seem not a lot of people do full SSD
setup.(Or share there experiences)
See this:
http://lists.opennebula.org/pipermail/ceph-users-ceph.com/2014-October/043949.htmlI'd suggest you try the above tests yourself, you seem to have a
significant amount of hardware already.
There are many SSD threads, but so far there's at best one example of a
setup going from Firefly to Giant and Hammer.
So for me it's hard to qualify and quantify the improvements Hammer brings
to SSD based clusters other than "better", maybe about 50%.
Which while significant, is obviously nowhere near the raw performance the
hardware would be capable of.
But then again, my guestimate is that aside from the significant code that
gets executed per Ceph IOP, any such Ceph IOP results in 5-10 real IOPs
down the line.
Christian
Anyway still brainstorm this so we can work on some POC. Will you guys
posted here. ---
Anthony Lévesque
On Apr 29, 2015, at 11:27 PM, Christian Balzer <chibi@xxxxxxx> wrote:
Hello,
On Wed, 29 Apr 2015 15:01:49 -0400 Anthony Levesque wrote:
We redid the test with 4MB Block Size (using the same command as
before but with 4MB for the BS) and we are getting better result from
all devices:
That's to be expected of course.
Intel DC S3500 120GB = 148 MB/s
Samsung Pro 128GB = 187 MB/s
Intel 520 120GB = 154 MB/s
Samsung EVO 1TB = 186 MB/s
Intel DC S3500 300GB = 250 MB/s
You will need to keep _both_ of these results in mind, the 4KB and 4MB
ones. For worst and best case scenarios.
And those dd tests are indicators, not a perfect replication of what
Ceph actually does.
Looking at your original results of 920MB/s over 96 1TB EVOs those SSDs
are thus capable of handling about 20MB/s combined journal/data
traffic. Ignoring any CPU limitations of course.
Unsurprisingly the best of both worlds in the SSDs you compared is the
Intel DC S3500.
And that is the slowest write and least endurable of the Intel DC SSDs.
To make a (rough) comparison between Intel and Samsung, the Intel DC
S3500 are comparable to the Samsung DC Evo ones and the Intel S3700 to
the Samsung DC Pro ones.
I have not tested the DC S3610 yet but I will be ordering some soon
Those will be (for journal purposes) the worst choice when it comes to
bandwidth, as they use higher density FLASH, thus less speed at the
same size.
They are however significantly more durable than the S3500 ones at
only a slightly higher price, thus making them good candidates for a
combined journal/data SSD. IF your expected write load fits the
endurance limits.
See:
http://ark.intel.com/compare/75682,86640,71914
Since previously we had the journal and OSD on the same SSD Im still
wondering if having the journal separate from the SSD (with a ration
of 1:3 or 1:4) will actually bring more Write speed. This is the
configuration I was thinking of if we separate the Journal from the
OSD:
Your speed will go up of course.
However it will not reach the fullest potential unless you put really
fast SSDs (or a PCIe, NVMe unit like this:
http://ark.intel.com/products/79624/Intel-SSD-DC-P3700-Series-400GB-12-Height-PCIe-3_0-20nm-MLC)
in there, see below.
÷Each OSD_Node÷
Dual E5-2620v2 with 64GB of RAM
Underpowered CPU when dealing with small write IOPS.
If all/most your writes are nicely coalesced by the RBD cache this may
not be a problem, but without knowing what your client VMs will do it's
impossible to predict.
-------------------
HBA 9207-8i #1
3x1TB Samsung 1TB for the Storage layer + 1 Intel S3610 200GB for the
Journal 3x1TB Samsung 1TB for the Storage layer + 1 Intel S3610 200GB
for the Journal -------------------
HBA 9207-8i #2
3x1TB Samsung 1TB for the Storage layer + 1 Intel S3610 200GB for the
Journal 3x1TB Samsung 1TB for the Storage layer + 1 Intel S3610 200GB
for the Journal -------------------
1x LSI RAID Card + 2x 120GB SSD (For OS)
2x 10GbE dual port
I suppose you already have that hardware except for the journal SSDs,
right?
I would have forgone extra OS SSDs and controllers and put the OS on
the journal SSDs in a nice RAID10.
There would be between 6-8 OSD Node like this to start the cluster.
My goal would be to max out at least 20 Gbps switch ports in writes
to a single OpenStack Compute node. (Im still not sure about the CPU
capacity)
Maxing out the port can come in many flavors/ways. Most of which are
not realistic scenarios, meaning that your VMs are more likely to run
out of IOPS before they run out of bandwidth.
But to achieve something like 2GB/s writes your SSDs and/or journal
SSDs will have to handle that speed.
4x 200GB DC S3610s at 230MB/s each for journal SSDs are clearly not
going to do that.
4x 400GB DC S3700s at 460MB/s each will come pretty close to that
2GB/s, however they will cost about the same as 2x P3700 mentioned
below. 16x 800GB DC3610s as combined journal/data SSDs will give you
about 4GB/s, though.
If you want to keep using your 1TB EVOs (after having verified what
their top speed as data only SSD is, but with 16 I suppose it will be
sufficient), use 16 of them and 2 of the 400GB DC P3700 Series cards
mentioned above.
If all of the above sounds either slow or expensive you're quite right,
the old adage of "fast, good, cheap" (good being endurance here) still
holds.
And once more, having a realistic idea of how much writes will happen
in that cluster will be crucial to make the right decision here.
Having to replace your cheap SSDs after a few months when DC level SSDs
at twice the price would have lasted 5 years is something to think
about.
Christian
As anyone testes a similar environment?
Anyway guys, lets me know what you think since we are still testing
this POC. ---
Anthony Lévesque
On Apr 25, 2015, at 11:46 PM, Christian Balzer <chibi@xxxxxxx> wrote:
Hello,
I think that the dd test isn't a 100% replica of what Ceph actually
does then.
My suspicion would be the 4k blocks, since when people test the
maximum bandwidth they do it with rados bench or other tools that
write the optimum sized "blocks" for Ceph, 4MB ones.
I currently have no unused DC S3700s to do a realistic comparison and
the DC S3500 I have aren't used in any Ceph environment.
When testing a 200GB DC S3700 that has specs of 35K write IOPS and
365MB/s sequential writes on mostly idle system (but on top of Ext4,
not the raw device) with a 4k dd dsync test run, atop and iostat show
a 70% SSD utilization, 30k IOPS and 70MB/s writes.
Which matches the specs perfectly.
If I do that test with 4MB blocks, the speed goes up to 330MB/s and
90% SSD utilization according to atop, again on par with the specs.
Lastly on existing Ceph clusters with DC S3700 SSDs as journals and
rados bench and its 4MB default size that pattern continues.
Smaller sizes with rados naturally (at least on my hardware and Ceph
version, Firefly) run into the limitations of Ceph long before they
hit the SSDs (nearly 100% busy cores, journals at 4-8%, OSD HDDs
anywhere from 50-100%).
Of course using the same dd test over all brands will still give you
a good comparison of the SSDs capabilities.
But translating that into actual Ceph journal performance is another
thing.
Christian
On Sat, 25 Apr 2015 18:32:30 +0200 (CEST) Alexandre DERUMIER wrote:
I'm able to reach around 20000-25000iops with 4k block with s3500
(with o_dsync) (so yes, around 80-100MB/S).
I'l bench new s3610 soon to compare.
----- Mail original -----
De: "Anthony Levesque" <alevesque@xxxxxxxxxx>
À: "Christian Balzer" <chibi@xxxxxxx>
Cc: "ceph-users" <ceph-users@xxxxxxxxxxxxxx>
Envoyé: Vendredi 24 Avril 2015 22:00:44
Objet: Re: Possible improvements for a slow write
speed (excluding independent SSD journals)
Hi Christian,
We tested some DC S3500 300GB using dd if=randfile of=/dev/sda bs=4k
count=100000 oflag=direct,dsync
we got 96 MB/s which is far from the 315 MB/s from the website.
Can I ask you or anyone on the mailing list how you are testing the
write speed for journals?
Thanks
---
Anthony Lévesque
GloboTech Communications
Phone: 1-514-907-0050 x 208
Toll Free: 1-(888)-GTCOMM1 x 208
Phone Urgency: 1-(514) 907-0047
1-(866)-500-1555
Fax: 1-(514)-907-0750
alevesque@xxxxxxxxxx
http://www.gtcomm.net
On Apr 23, 2015, at 9:05 PM, Christian Balzer < chibi@xxxxxxx >
wrote:
Hello,
On Thu, 23 Apr 2015 18:40:38 -0400 Anthony Levesque wrote:
BQ_BEGIN
To update you on the current test in our lab:
1.We tested the Samsung OSD in Recovery mode and the speed was able
to maxout 2x 10GbE port(transferring data at 2200+ MB/s during
recovery). So for normal write operation without O_DSYNC writes
Samsung drives seem ok.
2.We then tested a couple of different model of SSD we had in stock
with the following command:
dd if=randfile of=/dev/sda bs=4k count=100000 oflag=direct,dsync
This was from a blog written by Sebastien Han and I think should be
able to show how the drives would perform in O_DSYNC writes. For
people interested in some result of what we tested here they are:
Intel DC S3500 120GB = 114 MB/s
Samsung Pro 128GB = 2.4 MB/s
WD Black 1TB (HDD) = 409 KB/s
Intel 330 120GB = 105 MB/s
Intel 520 120GB = 9.4 MB/s
Intel 335 80GB = 9.4 MB/s
Samsung EVO 1TB = 2.5 MB/s
Intel 320 120GB = 78 MB/s
OCZ Revo Drive 240GB = 60.8 MB/s
4x Samsung EVO 1TB LSI RAID0 HW + BBU = 28.4 MB/s
No real surprises here, but a nice summary nonetheless.
You _really_ want to avoid consumer SSDs for journals and have a
good idea on how much data you'll write per day and how long you
expect your SSDs to last (the TBW/$ ratio).
BQ_BEGIN
Please let us know if the command we ran was not optimal to test
O_DSYNC writes
We order larger drive from Intel DC series to see if we could get
more than 200 MB/s per SSD. We will keep you posted on tests if
that interested you guys. We dint test multiple parallel test yet
(to simulate multiple journal on one SSD).
BQ_END
You can totally trust the numbers on Intel's site:
http://ark.intel.com/products/family/83425/Data-Center-SSDs
The S3500s are by far the slowest and have the lowest endurance.
Again, depending on your expected write level the S3610 or S3700
models are going to be a better fit regarding price/performance.
Especially when you consider that loosing a journal SSD will result
in several dead OSDs.
BQ_BEGIN
3.We remove the Journal from all Samsung OSD and put 2x Intel 330
120GB on all 6 Node to test. The overall speed we were getting from
the rados bench went from 1000 MB/s(approx.) to 450 MB/s which might
only be because the intel cannot do too much in term of journaling
(was tested at around 100 MB/s). It will be interesting to test with
bigger Intel DC S3500 drives(and more journals) per node to see if I
can back up to 1000MB/s or even surpass it.
We also wanted to test if the CPU could be a huge bottle neck so we
swap the Dual E5-2620v2 from node #6 and replace them with Dual
E5-2609v2(Which are much smaller in core and speed) and the 450 MB/s
we got from he rados bench went even lower to 180 MB/s.
BQ_END
You really don't have to swap CPUs around, monitor things with atop
or other tools to see where your bottlenecks are.
BQ_BEGIN
So Im wondering if the 1000MB/s we got when the Journal was shared
on the OSD SSD was not limited by the CPUs (even though the samsung
are not good for journals on the long run) and not just by the fact
Samsung SSD are bad in O_DSYNC writes(or maybe both). It is probable
that 16 SSD OSD per node in a full SSD cluster is too much and the
major bottleneck will be from the CPU.
BQ_END
That's what I kept saying. ^.^
BQ_BEGIN
4.Im wondering if we find good SSD for the journal and keep the
samsung for normal writes and read(We can saturate 20GbE easy with
read benchmark. We will test 40GbE soon) if the cluster will keep
healthy since Samsung seem to get burnt from O_DSYNC writes.
BQ_END
They will get burned, as in have their cells worn out by any and
all writes.
BQ_BEGIN
5.In term of HBA controller, did you guys have made any test for a
full SSD cluster or even just for SSD Journal.
BQ_END
If you have separate journals and OSDs, it often makes good sense to
have them on separate controllers as well.
It all depends on density of your setup and capabilities of the
controllers.
LSI HBAs in IT mode are a known and working entity.
Christian
--
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/
--
Christian Balzer Network/Systems Engineer
chibi@xxxxxxx
Global OnLine Japan/Fusion Communications
http://www.gol.com/
