On Thu, Dec 29, 2016 at 2:51 PM, Adam Goryachev
[off-list ref] wrote:
On 30/12/16 05:50, Doug Dumitru wrote:
quoted
Mr. Goryachev,
I find it easier to look at these numbers in terms of IOPS. You are
dealing with 17,500 IOPS vs 100 IOPS. This pretty much has to be the
"commit" from the drive. The benchmark is basically waiting for the
data to actually hit "recorded media" before the IO completes. The
"better" drive is returning this "write ACK" when the data is in RAM,
the the "worse" drive is returning this "write ACK" after the data is
somewhere much slower (probably flash). I would note that 17,500 IOPS
is a "good" but not "great" number.
So what would you consider a great number? I guess in practice the
environment isn't really that massive, it shouldn't really *need* great
numbers, but it seems no matter how hard I try to "over-architect", it is
still not performing to end user expectation.
The Intel drives run the same IOPS regardless of pre-conditioning.
They do this mostly by intentionally slowing down random writes so
that the worst case does not actually look any worse. You can pretty
much dial in any level of random write IOPS by manipulating over
provisioning. With 8% OP, a drive might get 5K IOPS, but at 20%, this
goes up to 15K. So if you want to keep an SSD fast, don't fill it up.
quoted
Doing commit writes to Flash is expensive. Not only do you have to
wait for the flash, but you have to update the mapping tables to get
to the data. Flash also does not typically allow 4K updates (even
given the erase rules), so your 4K sync update probably has to update
a 16K "page" is probably causing a lot of flash wear. Maybe as much
as 10:1 write amplification. Maybe more.
Wear doesn't seem to have been a problem so far.
9 Power_On_Hours_and_Msec 0x0032 000 000 000 Old_age Always
- 914339h+46m+34.180s
This is obviously wrong, I haven't had the drive for >100 years, but it is
at least almost 4 years old (early 2013 I suspect).
233 Media_Wearout_Indicator 0x0032 095 095 000 Old_age Always
- 0
This is the worst drive out of the whole array, the best is 99, but either
way it suggests these drives could easily last >10 years, which would be
well and truly longer than their expected/useful life (based on capacity).
241 Host_Writes_32MiB 0x0032 100 100 000 Old_age Always
- 3329773
This is the drive with the highest number of writes.... Obviously most
writes are smaller than 32MB, so I'm not entirely sure what this means, but
I suspect we are not doing a lot of writes per day compared to the total
storage capacity...
3329773 * 32MB / 3 years / 365 days = 97308MB/day. Total capacity is 480*7 =
3360GB or approx 0.03 per drive writes per day.
I've actually asked this question before, but here again we find what
appears to be an anomaly... some drives have significantly more writes than
others, and I don't understand why in a RAID5 array this would be the case,
I would have expected the writes to be split approx equally across all
drives...
225 Host_Writes_32MiB 0x0032 100 100 000 Old_age Always
- 1501762
225 Host_Writes_32MiB 0x0032 100 100 000 Old_age Always
- 1712480
225 Host_Writes_32MiB 0x0032 100 100 000 Old_age Always
- 1684811
225 Host_Writes_32MiB 0x0032 100 100 000 Old_age Always
- 1781849
225 Host_Writes_32MiB 0x0032 100 100 000 Old_age Always
- 2282764
225 Host_Writes_32MiB 0x0032 100 100 000 Old_age Always
- 2269957
225 Host_Writes_32MiB 0x0032 100 100 000 Old_age Always
- 2154155
225 Host_Writes_32MiB 0x0032 100 100 000 Old_age Always
- 2163563
225 Host_Writes_32MiB 0x0032 100 100 000 Old_age Always
- 3329774
It's hard to calculate whether some drives were replaced or similar due to
the nonsense power on hours values.... but generally all these drives were
purchased at the same time, and so should have been used mostly equally.
quoted
In the case of RAID-5 in a "traditional" Linux deployment, and
especially with DRBD protecting you on another node, you are probably
fine without having every last ACK "perfect". After all, if you power
fail the primary node, the secondary will become the primary, and any
"tail writes" that are missing will get re-sync'd by DRBD's hash
checks. And because the amount of data being re-synced is small, it
will happen very quickly and you might not even notice it.
Right, and at this stage I'm not even looking at data integrity, I'm only
examining "performance". In fact, it would be within the "acceptable"
parameters" to lose some data under a "disaster" scenario (where disaster
means losing both primary and secondary in an unclean shutdown). Of course,
I wouldn't design the system to do that, but it isn't a strict requirement,
as long as "normal" processes mean no data loss/corruption, and any drive
should (eventually) write all the data it has told you it will.
quoted
Back to performance, you should also consider what your array is doing
to you. You are running an 8 drive raid-5 array. This will limit
performance even more because every write becomes 2 sync writes, plus
6 reads. With q=1 latencies, if you run this test on the array with
"good" drives, you should probably get about 15K IOPS max, but it
might be a bit worse as the read and write latencies add for each OP.
Right, and one thing I've considered was moving to RAID10 to avoid this, but
even RAID10 means 2 writes. Assuming reads are relatively quick, than that
should reduce the impact of the RAID5 as well. At this stage, converting to
RAID10 is still something I'm holding up my sleeve as a last resort (due to
the additional wasted capacity).
Note that my tests are all on single drives, not the array. I can't afford
to be doing testing on the full array due to the destructive nature, and
also it is almost impossible to get a quiet moment where the tests wouldn't
be affected by workload.
quoted
I tried your test on our "in house" "server-side FTL" mapping layer on
8 drives raid-5. This is an E5-1650 v3 w/ an LSI 3008 SAS controller
and 8 Samsung 256GB 850 Pro SSDs. The array is "new" so it will slow
down somewhat as it fills. 439K IOPS is actually quite a bit under
the array's bandwidth, but at q=1, you end up benchmarking the
benchmark program. (at q=10, the array saturates the drives linear
performance at about 900K IOPS or 3518 MB/sec).
root@ubuntu-16-24-2:/usr/local/ess# fio --filename=/dev/mapper/ess-md0
--direct=1 --sync=1 --rw=write --bs=4k --iodepth=1 --runtime=60
--time_based --group_reporting --name=ess-raid5 --numjobs=1
Would it be possible for you to run the test on a single drive directly
instead?
quoted
Run status group 0 (all jobs):
WRITE: io=102653MB, aggrb=1710.9MB/s, minb=1710.9MB/s,
maxb=1710.9MB/s, mint=60001msec, maxt=60001msec
I might be looking at the wrong value, but you are getting 1711MB/s out of
an 8 drive array, I got a max of 99MB/s on a single drive, even if I
multiply that by 7 (8 drives - 1 redundancy), it's still less than half. I'd
be pretty keen to see your single drive results. Also whether those results
will change when using the 800GB model.
My test is of a "managed" array with a "host side Flash Translation
Layer". This means that software is linearizing the writes before
RAID-5 sees them. This is how the major "storage appliance" vendors
get really fast performance. One vendor, running an earlier version
of the software I am running here, was able to support 5000 ESXI VDI
clients from a single 2U storage server (with a lot of FC cards). The
boot storm took about 3 minutes to settle.
Single drives are around 500 MB/sec which is 125K IOPS through our
engine. Eight drives are (8-1)x500=3500 MB/sec or 900K IOPS. This is
actually faster than FIO can generate a test pattern from a single
job. It is also faster than stock RAID-5 can linearly write without
patches.
In terms of wear, lots of users are running very light write
environments. This is good as many configurations are > 50:1 write
amp if you measure "end to end". By end to end, I mean, how many
flash writes happen when you create a small file inside of a file
system. This leads to "file system write amp" x "raid write amp" x
"SSD write amp". Some people don't like this approach as the file
system is often "off limits" and a black box. Then again, some file
systems are better than others (for 10K sync creates, EXT4 and XFS are
both about 4.4:1 whereas ZFS is a lot worse). And with EXT4/XFS, you
can mitigate some of this with an SSD or mapping layer that compresses
blocks.
Doug Dumitru
Thank you for your advice, I'll see whether I can find a way to purchase one
of the samsung drives for testing/evaluation, then seem to be a similar
price to the Intel S3510 that I was looking at.
Regards,
Adam
quoted
On Wed, Dec 28, 2016 at 6:14 PM, Adam Goryachev
[off-list ref] wrote:
quoted
Apologies for my prematurely sent email (if it gets through), this one is
complete...
Hi all,
I've spent a number of years trying to build up a nice RAID array for my
SAN, but I seem to be slowly solving one bottle neck only to find another
one. Right now, I've identified the underlying SSD's as being a major
factor
in that performance issue.
I started with 5 x 480GB Intel 520s SSD's in a RAID5 array, and this
performed really well.
I added 1 x 480GB 530s SSD
I added 2 x 480GB 530s SSD
I now found out that performance of a 520s SSD is around 180 times faster
than a 530s SSD. I had to run many tests, but eventually I found the
right
things to test for (which matched my real life results), and the numbers
were nothing short of crazy.
Running each test 5 times and average the results...
520s: 70MB/s
530s: 0.4MB/s
OK, so before I could remove and test the 520s, I removed/tested one of
the
530s and saw the horrible performance, so I bought and tested a 540s and
found:
540s: 6.7MB/s
So, around 20 times better than the 530, so I replaced all the drives
with
the 540, but I still have worse performance than the original 5 x 520s
array.
Working with Intel, they swapped a 530s drive for a DC3510, and I then
found
the DC3510 was awesome:
DC3510: 99MB/s
Except, a few weeks back when I placed the order, I was told there is no
longer any stock of this drive, (I wanted 16 x 800GB model), and that the
replacement model is the DC3520. So I figure I won't just blindly buy the
DC3520 assuming it's performance will be similar to the previous model,
so I
buy 4 x 480GB DC3520 and start testing.
DC3520: 37MB/s
So, 1/3rd of a DC3510, but still better than the current live 540s
drives,
but also still half the original 520s drives.
Summary:
520s: 70217kB/s
530s: 391kB/s
540s: 6712kB/s
330s: 24kB/s
DC3510: 99313kB/s
DC3520: 37051kB/s
WD2TBCB: 475kB/s
* For comparison, I had a older Western Digital Black 2TB spare, and ran
the
same test on it. Got a better result than some of the SSD's which was
really
surprising, but it's certainly not an option.
FYI, the test I'm running is this:
fio --filename=/dev/sdb --direct=1 --sync=1 --rw=write --bs=4k
--iodepth=1
--runtime=60 --time_based --group_reporting --name=IntelDC3510_4kj1
--numjobs=1
All drives were tested on the same machine/SATA3 port (basic intel
desktop
motherboard), with nothing on the drive (no fs, no partition, nothing
trying
to access it, etc..).
In reality, I tested iodepth from 1..10, but in my use case, the
iodepth=1
matches is the relevant number. At higher iodepth, we see performance on
all
the drives improve, if interested, I can provide a full set of my
results/analysis.
So, my actual question... Can you suggest or have you tested any Intel
(or
other brand) SSD which has good performance (similar to the DC3510 or the
520s)? (I can't buy and test every single variant out there, my budget
doesn't go anywhere close to that).
It needs to be SATA, since I don't have enough PCIe slots to get the
needed
capacity (nor enough budget). I need around 8 x drives with around 6TB
capacity in RAID5.
FYI, my storage stack is like this:
8 x SSD's
mdadm - RAID5
LVM
DRBD
iSCSI
From my understanding, it is DRBD that makes everything a iodepth=1
issue.
It is possible to reach iodepth=2 if I have 2 x VM's both doing a lot of
IO
at the same time, but it usually a single VM performance that is too
limited.
Regards,
Adam
--
Adam Goryachev Website Managers www.websitemanagers.com.au
--
To unsubscribe from this list: send the line "unsubscribe linux-raid" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
--
Adam Goryachev Website Managers www.websitemanagers.com.au
--
To unsubscribe from this list: send the line "unsubscribe linux-raid" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
--
Doug Dumitru
EasyCo LLC