LSI 9211-8i

[XS Janus]

so I will be testing with 4x X25-E.

Game Over?

[Jor3lBR]

I felt like testing one of these myself and since it's so cheap I ordered one express from newegg... I could have it tomorrow if I am very lucky, but Canadian newegg kinda blows so I doubt it. Also ordered a 30gb agility to free up another X25-E that's being used by my laptop now, so I will be testing with 4x X25-E.

PCMark is influenced by cache so controllers with cache will probably do better.

PCMark05 with no tweaks. No cache results are allowed because you will not be able to submit that results to Futuremark. It has to be a valid OCB link. The Windows startup limit of 220 has to be respected, if cache is used this will go to the roof.

[Computurd]

this card is not benching well for the pcmarks. it is not scoring high enough because it does not have read ahead etc.

@Anvil Those results are positively beautiful man. i am wondering with your experiences with your lsi 9260-8i how would you compare that to last gen cards, in particular the 1231 areca?

I have been playing with the new card and i hate to put a damper on anything, but outside of using it as a dynamic disk i am not so sure it is faster in everyday usage as the 9260-8i. the 9211-8i is a beautiful card though lol,, but if you are using it as a boot card the 9260 can hit the same numbers without the hit in the sequential performance that this card is currently experiencing. maybe i am just used to seeing that big 1.7gb/s number and getting all hot and bothered by that LOL but performance wise i think there isnt much difference. okay the 9211-8i has a access time of .08 and the 9260-8i is around .14. BUT the 9260 can hit some SERIOUS 4k sequential and other sequential that the 9211 cannot. this is outstanding i/o of 64 on the 9260-8i at 4k sequential


NOW on the 9211 that 4k sequential is going to be equal to or less than the number of 4k random that it gets so basically around 290 mb/s when used as a boot disk.
so is the speed increase in sequential making up for the difference of the access time? and also, when under 'load' so to speak the access times are actually better than the 9211-8i. however so slightly that that may be... so here is a 4k random on the 9260-8i.


this is the 9211 with lower outstanding i/o. note the higher access time and the lower random results


very interesting results here and i plan to tailor a big time testing regimen between the two cards. i will definitely need a complete set of tests at different file sizes and outstanding i/o to test at.
I am thinking that on each card i should run:
4k Random & Seq
8k seq & random
32k seq & random
128k sequential and random
all should be ran first with 8 outstanding then 16 and 32 and 64. anything above that is pointless i would think above 64 outstanding i.o is not realistic for usage.

the jury is still out when it comes to 9260 v 9211. these cards are heavyweights!!! One caveat: the 9211 when used as a dynamic disk smokes anything bar none. and price v performance you would be crazy not to say the 9211 is more bang for the buck!

@one hertz- i am glad you are gonna give it a whirl as well, your input will be appreciated and you have great gear for testing

@spoiler cpu usage is higher, but slightly. nothing big tbh. i will be sure to make that a very big part of testing to help answer that question for you. i will also time the boot. i think my board does not like the controller lol cause it is hanging at the verifying dmi pool area and no other card ive had on this board has done that. and i have had like six different high performance cards on here ! However typical bios initialization for a raid card otherwise.
32 seconds for card initialization of its own bios.
strange dmi pool hang is 16 seconds.

[GullLars]

Yay! Finally a mod saw fit to authenticate me
Greetings to you all, I'm new as a participant here on XS, but i've been reading a lot of SSD-related threads and lurking on XS for almost 2 years now. I hang out with Nizzen and Anvil on a norwegian forum where we work on two major SSD-threads, where I initiated the first and biggest one a bit over a year ago, and the other one is new this fall and focuses on benchmarks. The thread is in norwegian, but the performance numbers, settings, and unit names should be understandable and is sumarized in the first post.
I saw you requested suggestions for benchmarking your 9260-8i vs 9211-8i and also compare to Areca 1231, so I thought I'd share and suggest.

I'd just like to apologize right away for a lot of text... :P

As i enter the thread, I am aware of the current avalible setups for testing:
Computurd: LSI 9260-8i and 9211-8i with 8 OCZ Vertex
Nizzen: Areca 1680ix with 4GB RAM, other misc RAID controllers, a bunch of Vertex (12?), and some x25-M. And 10(+?) 15K SAS.
Anvil: LSI 9260-8i, PERC 6/i, 4x x25-M 160 G2, 4x x25-M 80 G1, (+1 x25-M 160GB G2 in laptop and a retired 250GB Vertex).
One-Herz: Areca with unnamed SSDs. x25-E?
Tiltevros: LSI 9211 and 6 (5?) x25-M.
Biker: LSI 92xx with unknown # OCZ Agility.
And possibly:
Stevero: 3 Acards in 6x4GB R0 on areca 1231ML. (1231ML/8xR0acard ?)
Jor3lBR: Kingston ? and 4x Supertalent ? 32GB.

Sadly i won't be able to contribute with bench numbers at this time. I've only got a Adaptec 5805, 2x Mtron Pro 32GB (bought aug 08) and 2x OCZ Vertex 30GB. I haven't got ICH10R either, as I'm on a AM2+ Phenom 9850 setup. Not an extreme system per say :P

SO, regarding the suggestions for benchmarking:
I suggest 3 IOmeter setups each with a range of Queue Depth (QD) runs. For the controllers with cache, test lenght should be at least 3-4 times cache to avoid meassuring the cache speed instead of the RAID. Also, the controllers with cache should have a run time long enough to ensure any burst to/from the cache is insignificant.

IOmeter setup 1: 100% read 100% random 4KB, 5 sec ramp-up to avoid latency spike at start.

Setup 2: 100% write, 100% random 4KB, 5 sec ramp-up for controllers whitout cache and cache/[cache brust speed] sec for those with cache.

Setup 3: either 75% read or 66% read (what do you think?), 100% random, same ramp-up as setup 2.

Cache burst is determined with CDM 3.0 50 or 100MB lenght (wichever scores highest) and is 4KB 32QD numbers.

Test lenght 2GB whitout cache, 4x cache for controllers with cache. Runtime 30 sec for controllers whitout cache, 1 min for controllers with cache.

QD 1-256. Suggested QD stepping 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 12, 14, 16, 24, 32, 64, 128, 256. I suggest initially 1 worker, and we can eventually compare the difference queues of same total lenght being split on more workers makes. On setups with a high number of devices in RAID we could possibly change the steppings to reflect the total number of channels?

These 3 tests with 18 QD steps each will provide us with enough raw data to run a fairly complete analysis of controller and SSD strenght regarding random IO performance in terms of IOPS and accesstimes. I know this will take some time, but i think it will definetly be worth it. Maybe we should just start with setup 1 to validate?
As an example, here is Test 1 with almost the same parameters done on an x25-M connected with eSATA to a laptop. I know this is a bit different from what we will be testing, but the analysis method is the same and should bring usefull data and graphs. Benching data provided by Anvil, I've crunched the numbers.
Link to benchmark screenshots. (click spoilers to see screenshots)
Links to graphs generated from data:
IOPS by QD
Average accesstime by QD
Max accesstime by QD (this one went bad because of eSATA and craptop)
Snapshot of the spreadsheet used for first 3 graphs
And then the complicated and even more interresting stuff:
IOPS/average accesstime by QD
IOPS vs IOPS/accesstime by QD (2 competing graphs)
IOPS + IOPS/accesstime by QD (2 stacked graphs)
And finally link to excell 2007 spreadsheet with all raw data typed in, calculations (a few notes) and graphs. (XS wouldn't allow me to upload .xlsx, so it put it on a fileshare service)
Personally i love the last graph here with IOPS+IOPS/accesstime, as it depends on both high IOPS with simultaneously low accesstime for high scores. I bet ioDrive and ioXtreme will own at this particular graph, as they are designed for low latency and with massive parallell design.

I would love to hear what you think of this 4K random read analysis of x25-M over eSATA to craptop i got done yesterday

[SteveRo]

Welcome GullLars!
Wow you norwegians have been busy.
Thanks for all the good info.
I will execute setup 1 with 4 acard "boxes" each box set to R0 so that's 8xR0 on a 1231ML-2G.
Question for you, after ramp up, how long do you run the test for - 1 min, 2min? Also what should we use for a QD?

[Tiltevros]

Tiltevros: LSI 9260i4 or 9211i8 with 5 x25-m

[SteveRo]

Good morning again GullLars - below is 8xR0acard on 1231ML-2G - i7 cpu @ 4.48 ht off.
Do you want the results file?
The interesting numbers should be the new LSI cards.

The data file is located at -

http://dump.no/files/948f945c05e3/St..._100_read_.csv

[Ao1]

@ GullLars

[GullLars]

Thanks for the welcome guys
I'll se if i can get another norwegian guy here. The starter of our SSD bench thread is Ourasi, and he holds our records for single and 2xRAID-0 results in most benches.

I will execute setup 1 with 4 acard "boxes" each box set to R0 so that's 8xR0 on a 1231ML-2G.
Question for you, after ramp up, how long do you run the test for - 1 min, 2min? Also what should we use for a QD?

After ramp up (5-10sec), 1 min test run should do, since there are so many queue depths to test. Results really don't differ that much from 1 to 2 min.
I see you found the queue depths after you posted this.

Those are really good numbers with the Acard RAID, but you hit 90% of max IOPS at QD=5 on a 8x RAID-0, wich tells me your RAID is bottlenecked. Either that or the Acard units don't scale well IOPS wise when using 2 ports (2 ports don't mean 2x IOPS of 1 port). You have 28% of max IOPS at QD=1, so you get less than 4x IOPS scaling on this RAID.

It would be nice to get the excell file so i can make some graphs. At the moment i'm about to make some graphs of Anvils 250GB Vertex (1.40FW TRIM) he just posted IOmeter data for in our norwegian SSD-bench thread.
If you want I can link to the graphs here when i'm finished.

[SteveRo]

GullLars - Yes, I suspect I am saturating the 1231 - I will need a new controller
I will post the results spreadsheet if you have a location I can post to.

[GullLars]

I have a norwegian site that is excellent for quick sharing of small files:
http://dump.no/
Just click the line or browse, select your file, and click the green arrow. It uploads, and gives you a link to the file you can post here (or in a PM to me ).
Maybe you could just edit your post with the screenshot so the URL to the file is below it? That way anyone who wants can get the raw data.

[Tiltevros]

this is with 4 intels x-25m G1s and LSI 9260i4
4k read 100% random 100% with 64Q

[Tiltevros]

same but 100% seq.

[GullLars]

Stevero, do you mind testing a single Acard box on your areca on both one port and two ports? This will help establish if it can deliver more IOPS on two ports, how much scaling you got from more boxes on your last numbers, and how the Acard unit itself scales with QD.
I request QDs 1,2,3,4,5,6,7,8,16,32. Fine granularity in the start to see scaling pr QD, and then two big jumps to get a look if it has more to offer. Your last numbers had 80% of IOPS performance from 4 boxes using dual connection at QD=4, wich makes me suspicious about the scaling pr QD of single boxes. With SSDs you normally get almost perfect scaling untill QD=channels (a little controller overhead and possibility of requests to same channel in queue), and then a sudden dropoff. In RAID you should also get almost perfect scaling untill QD = units*channels, with only controller overhead and uneven distribution to units/channels in the queue reducing IOPS/QD scaling.

Tiltveros, that is apalling random IOPS scaling! A single x25-M can do 35 or even 40K (random) IOPS from ICH10R. Would you mind connecting 4x x25-M to ICH10R and doing a IOmeter run with 4KB 100% read 100% random QD=32?
I know that only makes 8 I/Os outstanding pr unit (wich have 10 channels each), but with x25-M's scaling a single unit get >25K IOPS at QD=8, so 4 at QD=32 should get >100K.
Anvil has even done 103K IOPS with 3 x25-M from ICH10R.

[SteveRo]

I have a norwegian site that is excellent for quick sharing of small files:
http://dump.no/
Just click the line or browse, select your file, and click the green arrow. It uploads, and gives you a link to the file you can post here (or in a PM to me ).
Maybe you could just edit your post with the screenshot so the URL to the file is below it? That way anyone who wants can get the raw data.

done, thanks for the link.

[SteveRo]

In playing around with the acards I discovered something interesting.
If you create a partition smaller than your cache you can check the read speed of just the cache - this with the 1231ML-2G, one acard and one stick of 2GB ram -

[Anvil]

A pity they removed QD4 from CDM 3, QD4 was included in the Technical Preview.

Nice results on read though

[SteveRo]

4KB 100% read 100% random - 1231ML-2G - one acard 1xR0
Output file at -

http://dump.no/files/948f945c05e3/12...100_random.csv

[SteveRo]

4KB 100% read 100% random - 1231ML-2G 2xR0 acard
Output file located at -

http://dump.no/files/948f945c05e3/12...100_random.csv

[SteveRo]

I can get higher small file IOPs on ich10r - this with 6xR0 acard -

Output file located here - http://dump.no/files/65cc710d3e0c/ic...6xR0_acard.csv

[Computurd]

i thought this was 9211-8i thread?
very nice results steve-ro that is impressive them acards SMOKE!!!

[One_Hertz]

I love newegg. I paid for express shipping and they shipped expedited

[Tiltevros]

update.... 7 intels x25ms from now on... with LSI 9211 filesize and profiles bench for iometer are welcome
i waiting to do ur task

[Computurd]

nice tiltervros! LOL that rocks. run some 4k randoms and show us the goods...however lets see some other stripe sizes as well, 4k is not the end all! i would just run 4k random at a few different queues , like 1 8 16 32 64 to start. i have the 9260 on tyhe bench right now playing around, gonna do some testing of game level load times etc and some other benchmarks to compare to the 9211-8i

@gullars/stevero did you ever get the graphs made for the 1231 and acards? that is gonna be cool!!

[Tiltevros]

512B 4K 8K 16K 32K random and seq. @ 32Q.
7x intels x25m @ 9211i8