LSI 9211-8i

[Spoiler]

Is it the one that's dated 12-NOV-99? That's the only one I see.

[Computurd]

yea, before they didnt even have a revision out. i will be installing array later today..

[fhj52]

It is just the SAS Flash utility & BIOS stuff that is from Oct & Nov. The FW, 2118ir.bin, is dated 23 Dec 2009.

Note that LSI stated to me in email

...

The firmware upgrade is not necessary.

[...snip...]
Technical Support Engineer
Global Support Services

LSI Corporation

I won't be upgrading until at least they tell me|us what the heck the changes are.

[Tiltevros]

how do we flash the bin file? i update the bios from the megaraid storage manager but i cant flash the FW... help anyone??

[Computurd]

its not easy LOL
flash bios from within MSM
you have to use a CLI to do the firmware, its the only method.
(i had to call support to get it to work)
take the contents of FW file and place in a generic folder on the desktop
then place the executable file (sas2flash.exe) from the sas2flash folder that corresponds with your OS in the SAME folder
open CMD prompt that is elevated to administrator status
inside commanbd prompt navigate to folder containing the FW and the sas2flash file
type in (sans quotes of course)
"sas2flash -f 2118ir.bin"
good luck any questions hit me on messsenger on kansas time!

[Tiltevros]

ok here is some results from the 9211i8 with the new flash...

the first and the second image is random and seq @ 4k file with queue of 1

the 3rd and the 4rth are 4k file with 32 queue

15,7MB more in seq and 9MB in random.. thats a quite good improvement

[Tiltevros]

now the big big big improvement for me is @ 16k files.... i dont know how they do it but i have 250-300 MB more..... in seq. files.... now my score in 32k files seq. is 1618,90MB/s from 1,3-1,4GB/s

now for 64k files my score is the same cause of the limiting of HDD so no problem in that..

[Computurd]

looks good..i havent had time to test yet, got os ready with programs installed...

[trans am]

NICE TILT paul get in the game man! I'm waiting.. I'm flashing the replacement 9260 i'll be back soon.

[Tiltevros]

paul beat this 1702 MB/s 64k @ 32que
0,0624 ms @ 4k file 1 queue

[trans am]

paul beat this 1702 MB/s 64k @ 32que
0,0624 ms @ 4k file 1 queue

Tilt, if you or paul can give me the steps to config iometer the same way you are so we're all on the same page I would appreciate it. I have never used it before this. I'm used to clicking crystal disk mark and getting a beer from the fridge and hitting alt/printscreen ctrl c/

i went and flashed the 9260 and it went so smoothly on the rampage ex. I being on that damn am3 crap since summer I forgot how computers are supposed to work.
I'm gonna call M.Koert at LSI tomorrow and ask him to join xs if he hasnt already. I think we should join forces with him directly to max these cards out and exchange feedback tips and board/drive compatibility. In exchange he can flow us some beta bios's to play with and some free hardware to test. Everyone wins. .. LSI gets top tier benchers squeezing every last drop of performance out of their hdwr and they say thanks and flow us some new toys. at least Paul(computurd) that man deserves some freebies. I'll vouch for you man.

on another note. do you remember those got milk commercials they did awhile back and show someone with a gigantic chocolate chip cookie and they run out of milk? I feel like that guy. I have my replacement 9260, a new i3 530 but the board, ssds and eco ram doesn't arrive till monday. I'm currently on a 1.6ghz pentium e2140 775 a rampage extreme running win7 64 on a single raptor. After that brief moment i had the ssds and lsi actually working last week compared to this crap I feel like I'm in the special olympics.

[Computurd]

@ trans am...well as to instructions for iometer YHPM...
tilt your performance with that card continues to excel way past mine...that rocks man, that latency is hard to match, as a matter of fact with my old trusty NF200 in the way it will be impossible right now at same QD with 4k it is .0855 so you are really kicking my arse there!

[Tiltevros]

http://www.xtremesystems.org/forums/...1&postcount=21

[trans am]

paul get back in here man its sat night? lets see that new fw bench

[Computurd]

thew new firmaware has done nothing to fix my issues with scaling and my array. it is very irritating. not to say that the performance isnt great, however it could be better. i am very demanding user

[GullLars]

I'm sorry i haven't been active here for a while, i have now read up on all posts and am up to speed.

I have a theory on why the IOPS performance hits a roof around 70-75K IOPS. It's probably the 800mhz powerPC controller that becomes the bottleneck from further IOPS scaling.
As to the large block sequential performance hit, i have no clue. It SHOULD be possible to fix in firmware, and i can't imagine it would be too hard.
Also, OCZ has added TRIM support to its newest Z-drive through custom drivers and firmware for the RAID controller. LSI should also be able to do this for the 92xx series, if they aren't confident enough to release it to enterprise users, they could at least release a beta version for us (the ones that have the controller anyway) to play with.
I support the suggestion of getting the guy from LSI to join the forum and use this thread as a sort of alpha/beta testing ground with feedback. I think LSI could benefit from using us, and it could speed up the development of good drivers and firmware for enthusiasts, wich will be a good market for the 92xx series.
I'm considering getting a 92xx controller myself, but i'm waiting a bit for Areca 18xx and the newer generation SSDs about to enter the market.

[Computurd]

it is actually a 533 mhz power pc..i had come to the same conclusion about the iops scaling, however tiltervros experiences and results with the intels has convinced me otherwise. i think the problem is with the 1.41 firmware i am using update is out today so we shall see, i am puttimng 1.5 on my drives soon.

[GullLars]

You're right. I forgot. I read a few articles on the 9260 wich has 800mhz.
One of the things that ought to make the 9211 cheaper is the lower frequency of the controller.

Still, if you look at it from a practical perspective, 70-75K 4KB random IOPS = 280-300MB/s. Anything even an EXTREME user would do in their wettest dreams would not be bottlenecked noticably by this. When you also take into account that the IOPS limit stays the same for larger blocks meaning 560-600MB/s for 8KB blocks, and 1120-1200MB/s for 16KB blocks, you can disregard this limitation of the controller for all practical scenarios.

What is more interresting is to check out the average accesstime and maximum accesstime for queue depths 1-128, or more related to practical scenarios QD 4-64, or even QD 8-32.

From a theoretical standpoint, an 9211 with 8 x25-M at QD 32 should be able to deliver accesstime comparable to 1 x25-M from ICH10R at QD 4. This will translate into what you call "CPU acceleration", causing the CPU to have less "wait" cycles while the blocks are fetched, and reducing experienced latency throughout the system.

This should be fairly easy to test out preliminary by benching 1 SSD from ICH10R at queue depths 1, 2, 3, 4, 8, 12, 16 and then benching 4 of the same SSD from 9211 (4i or 8i) at queue depths 4, 8, 12, 16, 32, 48, 64.
EDIT: Or bench the 1 SSD as non-member on 9211! Perhaps single test on both ICH10R AND 9211? Or even both single and 4xR0 on both controllers to look at accesstime and IOPS scaling?

What do you guys think?

[Tiltevros]

if u see from my results in LSI on TYAN m/b u will see 4k file sequential @ 1 QD is like 0,065ms - 0,064ms and i have 7 intels on 9211.
But yeah i will try to do the tests with 7 drives for fun

[GullLars]

Tiltevros, i'm talking about random accesstimes here, not seq. I hope you do the tests you said you will try with 4KB random access pattern. 4KB sequential at higher queue depths will likely have a lot lower accesstimes, but you won't see much of 4KB sequential in real scenarios.

[Computurd]

yes tilt lets see some results! i have test results from 4k on the 9211 that i have posted elsewhere, with eight vertex though....maybe you could give some insight into your idea from this gullars?

[trans am]

paul can you go to the 9260 thread quick and see me on gmail?

[GullLars]

Computurd, i'm confused if you mean what gave me the idea, or what i think we can do with the material.
I got the fundamental idea earlier when contemplating vertex vs x25-M IOPS scaling when i made the graphs i have linked here earlier of scaling with QD. Specifically when i made the the graphs average accesstime by QD, and IOPS/average accesstime by QD.
Basically, the more flash channels you have, the higher QD you can have before your accesstime starts to increase. Once you pass a certain number of channels and/or devices, the controller becomes really important to how much penalty you get by increasing QD even when QD < #channels.

The purpose of gathering this information is to see if various controllers are a good fit when the goal of the RAID array is to act as a CPU-accelerator (or system-accelerator) and not only to increase the storage throughput.
When looking at this, it is also relevant to compare the SSDs being used to HDD arrays and even RAM-SSDs. As far as i can tell, at low queue depths RAM-SSDs are best as they have the lowest accesstime, but once you pass #channels the latency doubles as you double QD. This means that while an ACARD SSD may own x25-M at QD=1 and QD=2 when used in dual-port mode, it doesn't have more channels and will double latency with queue depth, while x25-M has 10 channels and will scale much better with queue depth untill it overtakes the ACARD at average latency (and IOPS).

I suspect the LSI 9211 is good at delivering low latency IOPS in integrated RAIDs compared to other cards since it does not concern itself with caching or other fancy features. By testing the queue depths i suggested above, we can see how much latency-penalty (how far from perfect scaling) the controller gets by adding more channels and increasing the queue depth accordingly. We can also examinate how much is to gain at certain queue depths by increasing the number of units in the RAID.

If you test a single Vertex at queue depths 1, 2, 3, 4, 8, 16, 32 we have the corresponding data from your screenshot above if that is 8x Vertex R0.

[Tiltevros]

@ gulllars ----> http://www.xtremesystems.org/forums/...d.php?t=243725

[Computurd]

well i dont think there is a need to run those numbers on a single ssd. there isnt a ssd in the world that will get even close to those numbers^^^^