Interesting Windows Server 2012 LSI SAS2308 Controller Results

Notice: Page may contain affiliate links for which we may earn a small commission through services like Amazon Affiliates or Skimlinks.

dba

Moderator
Feb 20, 2012
1,477
184
63
San Francisco Bay Area, California, USA
If it would help, I could spin up a 9207 card with a Samsung 830 128GB on Win2008R2 and see if I can replicate the issue. Let me know if that would be useful to you or not.

Just tried AS SSD after installing .NET framework 3.5. Same result on the Samsung 830 and 840 Pro. 2.2MB/s 4K writes.

Also... I tried a Vertex 1 128GB. 12MB/s 4K writes on the LSI controller. Wonder if it is something Samsung related?
 

Patrick

Administrator
Staff member
Dec 21, 2010
12,511
5,792
113
if you had a single drive that would be awesome! I am thinking of installing win2k8r2 on it just to see.

Really cool to find something like this.
 

dba

Moderator
Feb 20, 2012
1,477
184
63
San Francisco Bay Area, California, USA
I'll give it a quick try - probably Monday (tomorrow). Test will look like this: To exising AMD Opteron server running Win2008R2, add a single Samsung 830 128GB drive attached to an otherwise empty LSI 9207-8e card. Benchmark using Anvil and, if the results look as bad as yours, using IOMeter as well. Will that work?
 
Last edited:

Patrick

Administrator
Staff member
Dec 21, 2010
12,511
5,792
113
Should work...

I do have some news. I installed Windows 7 x64 Ultimate on the machine. Seeing very similar results. Did you see anything strange with the Supermicro X9DR7-LN4F's onboard ports and the Samsung drives?

BTW: This looks similar. Not great information on the fix.

Update: Just checked, it is showing as 6.0gbps negotiated link speed.
 

mobilenvidia

Moderator
Sep 25, 2011
1,956
212
63
New Zealand
This is going to be interesting to see what happens.

Some things to consider:
Anything is PCIe slots 1,2 or 6 these are linked to CPU1 which the SAS2308 also hangs from.
Check AIDA if the SAS2308 is @ PCIe 3.0 8x
What is ATTO getting with the 840 pro ?
Factory reset on the System BIOS (or a performance setting if SuperMicro does this)
Did you reflash IT FW with no BIOS ?
Does SAS2flsh have a factory reset like Megacli does ? (there is a very good post here somewhere :D)

Beginning to think the LSI Firmware is having issues with the chipset of the SSD's.
Even the Vertex4 should be doing better.
 

Patrick

Administrator
Staff member
Dec 21, 2010
12,511
5,792
113
Ran IOmeter and played with a bunch of settings. I think this was a check-box issue.
 

dba

Moderator
Feb 20, 2012
1,477
184
63
San Francisco Bay Area, California, USA
We'll here is a first bit of data. When run with the Anvil default settings, I am, like Patrick, seeting extremely low write rates - less than 200 IOPS for 4kb writes regardless of queue depth. The fix is to use the Anvil's settings to "enable write-through". With this change, results look normal - 8,200 4kb write IOPS rising to 48,000 at a queue depth of four and 55,600 at queue depth 16. Also, the Anvil's tests are very short. I later changed the settings to make each thread run for 30 seconds, which I contend is a better real world test. This change also imnproved most scores.

I'll give it a quick try - probably Monday (tomorrow). Test will look like this: To exising AMD Opteron server running Win2008R2, add a single Samsung 830 128GB drive attached to an otherwise empty LSI 9207-8e card. Benchmark using Anvil and, if the results look as bad as yours, using IOMeter as well. Will that work?
 
Last edited:

Patrick

Administrator
Staff member
Dec 21, 2010
12,511
5,792
113
We'll here is a first bit of data. When run with the Anvil default settings, I am, like Patrick, seeting extremely low write rates - less than 200 IOPS for 4kb writes regardless of queue depth. The fix is to use the Anvil's settings to "enable write-through". With this change, results look normal - 8,200 4kb write IOPS rising to 48,000 at a queue depth of four and 55,600 at queue depth 16. Also, the Anvil's tests are very short. I later changed the settings to make each thread run for 30 seconds, which I contend is a better real world test. This change also imnproved most scores.
That was the check box. Looks like I have a filled 830 that is not recovering well, but results look fairly normal after checking the box. One very interesting thing is that Anvil's results seem to track 4K sequential writes on IOMeter not random.

If we have multiple folks seeing this, then I might do a quick main site post.