Post Reply 
 
Thread Rating:
  • 0 Votes - 0 Average
  • 1
  • 2
  • 3
  • 4
  • 5
Does SSD cache really speed up IO access?
05-04-2016, 08:32 AM
Post: #21
RE: Does SSD cache really speed up IO access?
(09-23-2015 06:37 PM)kirbyzhou Wrote:  In my opinion, the filesystem metadata is the most valuable things to cache, so the du speed can very directly to show you the problem.
But as your wish, I have done another mini bench to demonstrate the problem.
It is a little hard, because I need to avoid the OS level disk cache in memory.
If data is cached by OS, the IO would not passed to Drobo, but we need to check whether the cache inside Drobo is OK.
So I take a syscall to turns OS data caching off.

<snip>


Sorry for the delay in coming back on this - it's taken a while to find time to dig into things properly.

I'm running some more tests, and there's a couple of interesting things that have come up:

- On my 5D running firmware 3.5.0 and Dashboard 2.8.1, I'm getting comparable read and write figures of ~200-250MB/s from Blackmagic SpeedTest. I understand this software tests uncompressible data - but I'm guessing from the fact that it's video-focussed that it's sequential access rather than random access. However, it's interesting how the figures seem so very high...

- (More interestingly) I've just been running the Minibench test - thanks kirbyzhou - that's a really useful little utility you've concocted there. My initial readings were similar to yours:

disk1
KB/t tps MB/s
64.00 77 4.81
64.00 74 4.60
64.00 72 4.49
64.00 69 4.30
64.00 75 4.67
64.00 72 4.50
64.00 69 4.29
64.00 73 4.54
64.00 71 4.43
64.00 69 4.30
64.00 68 4.23
64.00 81 5.05
64.00 68 4.23
64.00 72 4.49
64.00 76 4.74
64.00 66 4.12
64.00 76 4.73
64.00 72 4.49
64.00 78 4.85
64.00 80 4.99 etc

But after some time (30 mins? 15 iterations?), the Drobo did some gurgling, and things slowed down for a short period:

disk1
KB/t tps MB/s
64.00 75 4.69
64.00 69 4.31
64.00 77 4.79
64.00 63 3.93
64.00 67 4.17
64.00 6 0.37
64.00 4 0.25
64.00 4 0.25
64.00 5 0.31
64.00 21 1.31
64.00 69 4.31
64.00 78 4.85
64.00 79 4.92
64.00 76 4.73
64.00 57 3.55
64.00 7 0.44
64.00 5 0.31
64.00 7 0.44
64.00 18 1.12
64.00 15 0.93

...And then things got noticably faster:

disk1
KB/t tps MB/s
64.00 112 7.03
64.00 111 6.94
64.00 111 6.92
64.00 113 7.06
64.00 110 6.90
64.00 117 7.29
64.00 112 6.98
64.00 112 7.03
64.00 112 6.99
64.00 112 7.00
64.00 112 6.98
64.00 110 6.85
64.00 113 7.04
64.00 104 6.48
64.00 111 6.91
64.00 109 6.78
64.00 109 6.80
64.00 114 7.12
64.00 109 6.81
64.00 114 7.10



Here's pretty chart showing the transfer rate from this process: https://www.dropbox.com/s/1hu2qhodqpvdpf...https://www.dropbox.com/s/1hu2qhodqpvdpfx/Drobo%205D%20Minibench%20w%20SS


I haven't yet had chance to reproduce the test or look further into it, but I was _very_ interested to see that something suddenly changed in its behaviour.

Is anyone else able to reproduce this?
Find all posts by this user
Quote this message in a reply
05-05-2016, 07:38 PM (This post was last modified: 05-05-2016 07:38 PM by Paul.)
Post: #22
RE: Does SSD cache really speed up IO access?
Drobo AI - "it's Aliiiive", and it adapts to your usage, adding to its technological distinctivenes Smile
also known as DroBorg Big Grin

(btw i have XP home SP2, a Drobo v1 with 2x 1TB/2x 1.5TB WD greens, & a bkp Drobo v2 with the same + a DroboShare: unused)
& a DroboS v2 with 3xWD15EADS &2x1TB in DDR mode on win7, & a drobo5D (all usb)
  • btw i did a sustained (write) operation for about 6 hours, and got 13.2MB / sec ...objection? "sustained" :)
    (16.7MB/s on a v2 & 47-96MB/s drobo-s)
Find all posts by this user
Quote this message in a reply
05-09-2016, 03:38 AM (This post was last modified: 05-09-2016 03:40 AM by MoisiePants.)
Post: #23
RE: Does SSD cache really speed up IO access?
Hello again:

Ok - I've been doing some more tests. All quite interesting...

So far I have been unable to reproduce the increase in performance that seemed to come about mid-test. However, here's what I have found:

Test 2 - comparing the first test with a 2.5" hybrid drive (HDD with small SSD caching - similar in principle to Drobo's SSD caching capability) in a USB 3.0 enclosure: the hybrid drive (over USB) has roughly the same performance as the Drobo (over TB), but does not increase at the end of the run:

https://www.dropbox.com/s/gykncpv37v8qwu...https://www.dropbox.com/s/gykncpv37v8qwuq/Test%202%20-%20Drobo%205D%20w%20SSD%20vs%20Hybrid%20Driv

Test 3 - comparing the first test exactly as is, but with the mSATA SSD removed: the performance is actually significantly improved for having the SSD not present:

https://www.dropbox.com/s/dx0f8ik83pzdzt...https://www.dropbox.com/s/dx0f8ik83pzdztv/Test%203%20-%20Drobo%205D%20w%20SSD%20vs%20Drobo%205D%20w%3Ao%20SS

Test 4 - comparing the first test with the mSATA SSD removed again, but having initialised the disk pack with Dashboard 2.8.1 and firmware 3.5.0: the performance is further improved over Test 3 (note I had very little data on the disk pack in either scenario, so it's not just comparing a really old crufty directory with a new one):

´╗┐https://www.dropbox.com/s/mx20wnxr9puf8yg/Test%204%20-%20Drobo%205D%20w%20SSD%20vs%20Drobo%205D%20%283.5.0%29%20w%3Ao%20SSD.pdf?dl=0

Test 5 - comparing the first test with the same setup as Test 4, but with the mSATA SSD reinserted - i.e. the only difference is that Test 5 had the disk pack initialised with the latest software. The performance is near-identical to Test 4 (no SSD):

https://www.dropbox.com/s/59e30p3hcz804a...https://www.dropbox.com/s/59e30p3hcz804ah/Test%205%20-%20Drobo%205D%20w%20SSD%20vs%20Drobo%205D%20%283.5.0%29%20w%20SS

And a composite of tests 1-5:

https://www.dropbox.com/s/9yevudmtqqs9ev...https://www.dropbox.com/s/9yevudmtqqs9evs/Tests%201-



I've got more tests to run, but it does so far look like:
- Having the mSATA SSD is not necessarily of benefit;
- Reinitialising the disk pack with the latest software does improve performance.


Will keep you posted with more results...
(09-27-2015 07:10 AM)kirbyzhou Wrote:  I have read your post.
I see that, you reported the transfer speed is around 300MB/s without SSD.
In my opinion, the top speed Drobo can fetch never exceed 300MB/s, so your benchmark is not meaningful.
SSD improve the IO performance mostly by saving the seeking time, not by increase the pure transfer speed.
I suggest you redesign your test. At first you must let your transfer speed become low than 20MB/s without SSD -- that must be caused by a lot of disk seeking.

Hi Kirbyzhou:

Sorry - yes, I think I was falling foul of memory caching, which obviously screwed up my tests! Apologies for the misinformation!
Find all posts by this user
Quote this message in a reply
05-09-2016, 09:39 AM
Post: #24
RE: Does SSD cache really speed up IO access?
If you are looking for "Fast", judging from my experience with the Drobo b810n (2x SSD-240's and 6x 4TB Constellations), you are not going to find it in a Drobo. Maybe the 1200i would provide a different experience but I don't think anything below that will. I spent an enormous amount of time getting mine into production and now I am going to go back to a Linux server for my storage and just co-lo the Drobo for offsite backups.
Find all posts by this user
Quote this message in a reply
05-13-2016, 06:17 AM
Post: #25
RE: Does SSD cache really speed up IO access?
Hello again:

Ok - I think I have some definite conclusions here: it's all to do with transfer size.

I was poking around at the Drobo with Iometer under Windows (trying to replicate the scenario that Drobo demonstrated with the B1200i in their YouTube video); my first attempt yielded something that looked as though the SSD was having a tremendous effect - a clear and repeatable difference between running with the SSD present and without it.

Looking for differences in the minibench tests vs the Iometer test, I saw that I was running Iometer with a 4kB transfer size - whereas minibench ran with 64kB transfers.

On changing Iometer to 64kB, I then saw results which mirrored the ones I saw in minibench. Here's a graph showing the four situations (I'm showing IOPs instead of transfer rate here):

https://www.dropbox.com/s/l5csj4c392lb56...https://www.dropbox.com/s/l5csj4c392lb56z/Iometer%20Test

The caching effect seems to be effective up to about 32kB; once I change the transfer size from 32kB to 48kB, the results immediately show the same pattern as the 64kB transfers.

Moving back to minibench, I lowered the transfer size to 4kB - and again witnessed the same effect here: the SSD is clearly having an effect. Here's a graph showing the first five tests (from my previous post) plus this one (again, I'm showing IOPs rather than transfer rate):

https://www.dropbox.com/s/uxj5h6yzdnv7i1...https://www.dropbox.com/s/uxj5h6yzdnv7i1i/Tests%201-

I've looked at the transfer sizes for my application usage, and they appear to be in the 56kB+ range - so it looks like I would never see any benefit from the SSD caching with the present software. :-( It would also explain why some people are seeing real-world benefits where others are not - it depends on how large the transfers are in a particular piece of software.


Perhaps we can try to persuade Drobo to widen the scope of the caching algorithms - I can't see what downside that would have for the unit...
Find all posts by this user
Quote this message in a reply
05-13-2016, 07:27 AM (This post was last modified: 05-13-2016 07:32 AM by AzDragonLord.)
Post: #26
RE: Does SSD cache really speed up IO access?
Thank you for all your efforts!

So are you saying that if the files being read are larger than 32k, then there's no real benefit to having a cache? For years I've run a 128 GB mSATA cache in my 5N and it still shows 100% life, which suggests it hasn't been touched. About 95% of what I store are video files from 2-40 GB in size.

Drobo (Gen2)
Drobo 5N
=============================
** Praise the Gods, Do Good, Act Bravely. **
Find all posts by this user
Quote this message in a reply
05-13-2016, 10:21 AM
Post: #27
RE: Does SSD cache really speed up IO access?
Hello AzDragonLord:

It's not quite as you're asking:

(05-13-2016 07:27 AM)AzDragonLord Wrote:  So are you saying that if the files being read are larger than 32k, then there's no real benefit to having a cache?

It's not the size of the files themselves, but the size of the chunks in which software is reading them. So, for playing back a video file for example, the computer will not have sufficient resources to load the entire video into memory, and then play it out. Instead it loads (for example) 00:00 -> 00:10 from disk into memory, and then starts playing that. Then, whilst it's playing, it will load 00:10 -> 00:20, and so on.

It the size of these chunks that determines whether the mSATA caches them; generally as an end-user, you won't have any control over how the software fetches the chunks off disk - it's in the hands of the software developer.
Find all posts by this user
Quote this message in a reply
05-17-2016, 06:47 PM
Post: #28
RE: Does SSD cache really speed up IO access?
(05-13-2016 06:17 AM)MoisiePants Wrote:  Hello again:

Ok - I think I have some definite conclusions here: it's all to do with transfer size.

I was poking around at the Drobo with Iometer under Windows (trying to replicate the scenario that Drobo demonstrated with the B1200i in their YouTube video); my first attempt yielded something that looked as though the SSD was having a tremendous effect - a clear and repeatable difference between running with the SSD present and without it.

Looking for differences in the minibench tests vs the Iometer test, I saw that I was running Iometer with a 4kB transfer size - whereas minibench ran with 64kB transfers.

On changing Iometer to 64kB, I then saw results which mirrored the ones I saw in minibench. Here's a graph showing the four situations (I'm showing IOPs instead of transfer rate here):

https://www.dropbox.com/s/l5csj4c392lb56...https://www.dropbox.com/s/l5csj4c392lb56z/Iometer%20Test

The caching effect seems to be effective up to about 32kB; once I change the transfer size from 32kB to 48kB, the results immediately show the same pattern as the 64kB transfers.

Moving back to minibench, I lowered the transfer size to 4kB - and again witnessed the same effect here: the SSD is clearly having an effect. Here's a graph showing the first five tests (from my previous post) plus this one (again, I'm showing IOPs rather than transfer rate):

https://www.dropbox.com/s/uxj5h6yzdnv7i1...https://www.dropbox.com/s/uxj5h6yzdnv7i1i/Tests%201-

I've looked at the transfer sizes for my application usage, and they appear to be in the 56kB+ range - so it looks like I would never see any benefit from the SSD caching with the present software. :-( It would also explain why some people are seeing real-world benefits where others are not - it depends on how large the transfers are in a particular piece of software.


Perhaps we can try to persuade Drobo to widen the scope of the caching algorithms - I can't see what downside that would have for the unit...

It seems interesting, I will write a mini bench under Mac OS X to verify it.
Find all posts by this user
Quote this message in a reply
10-30-2016, 08:32 PM
Post: #29
RE: Does SSD cache really speed up IO access?
(05-17-2016 06:47 PM)kirbyzhou Wrote:  It seems interesting, I will write a mini bench under Mac OS X to verify it.

This raised a question for me, is the mSATA SSD card only a read cache? I have a different application (Windows Media Center) where certain blocks of disk (the live TV buffer) are written over and over. This behavior doesn't seem to show good performance on my Gen 3. Would a 5D/Dt be better? SSD cache for frequently written blocks doesn't seem like a super great idea, what with write lifetime limits in SSD.
Find all posts by this user
Quote this message in a reply
10-31-2016, 08:38 PM
Post: #30
RE: Does SSD cache really speed up IO access?
hi rasunders i think it is read only for data integrity reasons

(btw i have XP home SP2, a Drobo v1 with 2x 1TB/2x 1.5TB WD greens, & a bkp Drobo v2 with the same + a DroboShare: unused)
& a DroboS v2 with 3xWD15EADS &2x1TB in DDR mode on win7, & a drobo5D (all usb)
  • btw i did a sustained (write) operation for about 6 hours, and got 13.2MB / sec ...objection? "sustained" :)
    (16.7MB/s on a v2 & 47-96MB/s drobo-s)
Find all posts by this user
Quote this message in a reply
Post Reply 


Forum Jump: