LSI 9211-8i versus LSI 9260-8i versus Areca 1231ML-2G

[SteveRo]

OK so here we go, I will have all three controllers and 12 days before I go back to work.

New vs old, faster h/w with still being worked out s/w vs old h/w with mature s/w - should be fun.

I plan on complete iometer, pcmark vantage HDD test and maybe everest testing (if needed) with four acard 9010s and maybe with some SLC mtron mobi's - if it looks like that would help clarify anything.

I will do all testing with W7x64 (beta 7100).
I haven't gotten around to buying W7 for my test machine yet.

So - please post back to me - what driver, f/w and controller settings should I use for each controller to gain maximum performance - both sequential and random?

For the 9211-8i - I recommend driver - 20270 dated 20Feb09 - driver that shipped with W7 beta bld 7100 (I think). F/W is what shipped on the controller - 2.00.50.00-IR, there is no new firmware (yet) that I can find. Controller settings are set to default as I have yet to figure out how to change any settings (any settings that make any difference in performance that is).

For the 1231ML-2G - For firmware - 1.47, dated 2009/8/19 (latest on the areca website).
For driver - I have been using storport driver 6.20.0.16 dated 8/19/2008. Is there a better driver? Anyone using the scsiport driver?
Tell me what you are using and what performance you are seeing.
On the areca website: The latest storport driver is v19 (I am using v16 up to now). The latest scsiport driver is also v19.
I have read that storport is supposedly newer than scsiport but I don't remember testing to see which might be faster.

For the 9260-8i - you tell me - what is best controller settings, f/w and driver!

_______________________________________________

RESULTS -

Comparison results summary (links to the three way comparison - in this thread) –

Pcmark vantage HDD test for the LSI 9211-8i, Areca 1231ML-2G, LSI 9260-8i and intel ich10R (onboard raid) –

http://www.xtremesystems.org/forums/...&postcount=114

The vantage pcmark suite results –

http://www.xtremesystems.org/forums/...&postcount=116

Game level load times by controller -

http://www.xtremesystems.org/forums/...&postcount=229

Iometer results for the controllers –

http://www.xtremesystems.org/forums/...&postcount=235

Last but not least – FYI - AS SSD and CDM for all three controllers –

http://www.xtremesystems.org/forums/...&postcount=240

That is all for now.

[One_Hertz]

Do you have any games at all?

[SteveRo]

Do you have any games at all?

You really want game load times, eh?
COD4, CrWarHead and FC2 - loading on the acards can be a space problem though.

[trans am]

Looking good steve. What are you doing exactly? Raid0?

[SteveRo]

8xR0 acard 9010s, (DDR2 Ram drives)

[trans am]

8xR0 acard 9010s, (DDR2 Ram drives)

of course! http://www.mwave.com/mwave/SKUSearch...iteria=AA74662

Guys while we're waiting for steve, I found a nice review on xbit for the LSI 9260 8i
http://www.xbitlabs.com/articles/sto...as9260-8i.html
http://www.xbitlabs.com/articles/sto...9260-8i_4.html

older roundup but still a good read and very informative.
http://www.xbitlabs.com/articles/sto...s-roundup.html

[One_Hertz]

You really want game load times, eh?
COD4, CrWarHead and FC2 - loading on the acards can be a space problem though.

That's great. Do you also have ICH10R handy?

[SteveRo]

Trans - yes, but not the B version, the straight 9010 (no B) can do R0 (each box) - http://www.acard.com.tw/english/fb01...&type1_idno=13
So via jumpers you can set each box to R0.
So if you have four boxes - each set to R0 - you have 8 drives available - for building an array.
Each array has 4 DDR2 slots available - I will be using 16 2GB sticks, 4GB in each "drive".
If money is no object you can use 8GB sticks for 64GB per box (unformatted) - I have not tested this capacity- myself.

[lowfat]

You really want game load times, eh?
COD4, CrWarHead and FC2 - loading on the acards can be a space problem though.

It is the only thing that matters

[SteveRo]

That's great. Do you also have ICH10R handy?

yep - x58 mobo.

[Serra]

I think the fairest test would be to use the latest available drivers for both on the latest available firmware with 64k stripe sizes. Sure, you can optimize more, but those are the default settings that 99% of people are going to use. People will complain that you weren't maximizing performance for benchmark X, but people will complain if you do as well and I think the only way to win is to level the playing field to defaults.

Now if you wanted to do some tweaking after getting that baseline down, that's another thing... but I would really consider that icing on the cake.

[XS Janus]

This will be fun!

[SteveRo]

I think the fairest test would be to use the latest available drivers for both on the latest available firmware with 64k stripe sizes. Sure, you can optimize more, but those are the default settings that 99% of people are going to use. People will complain that you weren't maximizing performance for benchmark X, but people will complain if you do as well and I think the only way to win is to level the playing field to defaults.

Now if you wanted to do some tweaking after getting that baseline down, that's another thing... but I would really consider that icing on the cake.

Generally I concur, but I have already found a problem with the latest driver for the 9211 - see post - http://www.xtremesystems.org/forums/...3&postcount=97

In this case I want to use the old driver unless someone finds something better.

[Computurd]

excellent this is gonna be awesome the prospect of putting them head to head is exciting! for the 9260 the best for ssd is adaptive read ahead, disk cache enabled, direct I/O, NCQ disabled on the card, always write back, and the highest pcie packet size that your motherboard will allow. pci-e should be clocked to the same for each card. i would suggest 105. the size of any everest test files should be large enough to dissallow cache runs, and the profiles and test sizes of files should be the same. so if you come up with some standard tests that would be great. It is going to be interesting how the various cards handle the sheer power of those acards...i agree a like stripe size should definitely be used so that we can get a good comparison of that. game loads times will be a must, but only for comparison on your system only, as the load times will be specific to your setup.
also pcmark vantage scores should be a big must as well. i think that benchmark is about the best out there for real world usage.
it is going to be hard to quantify which is actually faster in real world usage, as all three are fast as hell also i think you should try to use the crysis warhead as a a good indicator because if it is like the original crysis it has a built in load level timer so that we get a computer score that is not subject to error. stopwatch is kinda...i dunno we are dealing with milliseconds and such! i hope the other games have built in level load timers?
this is going to be a very fun amount of testing for you! if done right this is gonna be great!

EDIT: i agree with your plans for the 9211 driver, i have noticed the same performance issues with the new driver now that you pointed it out. the newest firmware/driver combo from lsi for the 9260 will be your best bet!

[SteveRo]

excellent this is gonna be awesome the prospect of putting them head to head is exciting! for the 9260 the best for ssd is adaptive read ahead, disk cache enabled, direct I/O, NCQ disabled on the card, always write back, and the highest pcie packet size that your motherboard will allow. pci-e should be clocked to the same for each card. i would suggest 105. the size of any everest test files should be large enough to dissallow cache runs, and the profiles and test sizes of files should be the same. so if you come up with some standard tests that would be great. It is going to be interesting how the various cards handle the sheer power of those acards...i agree a like stripe size should definitely be used so that we can get a good comparison of that. game loads times will be a must, but only for comparison on your system only, as the load times will be specific to your setup.
also pcmark vantage scores should be a big must as well. i think that benchmark is about the best out there for real world usage.
it is going to be hard to quantify which is actually faster in real world usage, as all three are fast as hell also i think you should try to use the crysis warhead as a a good indicator because if it is like the original crysis it has a built in load level timer so that we get a computer score that is not subject to error. stopwatch is kinda...i dunno we are dealing with milliseconds and such! i hope the other games have built in level load timers?
this is going to be a very fun amount of testing for you! if done right this is gonna be great!

EDIT: i agree with your plans for the 9211 driver, i have noticed the same performance issues with the new driver now that you pointed it out. the newest firmware/driver combo from lsi for the 9260 will be your best bet!

ok - very good, one comment and one question.
comment - my ggbt x58 extreme will only oc the pcie to 102 -so that is all I will be able to do
question - I thought you said 1MB or 512 K was best blocksize for 9260, I want to use what is best?

[Computurd]

well i agree that 1mb is best, however for comparison i agree that we should use the same, but then again do you think that matters? i just want to ensure a fair test.
i am also very curious of small file sequential with the 1231 as small file sequentials are very important as well in certain usage scenarios. i know this may be a bit much to ask but this is what i did with the 9260 a few days ago for some comparison to the 9211. maybe with a few of the QD would be fine like 1, 4, 8 , 16, 32 and 64 should be enough

RE: post 18

[trans am]

ok - very good, one comment and one question.
comment - my ggbt x58 extreme will only oc the pcie to 102 -so that is all I will be able to do
question - I thought you said 1MB or 512 K was best blocksize for 9260, I want to use what is best?

does this help?

http://www.xbitlabs.com/discussion/5878.html

You must log in to reply
2.
What is the reason for using the 64kb default stripe size?

In my (limited) experience using this card so far, I found far better results using a stripe size of 256k or 512k.

I currently have an 8 disk SATA Raid-0 array running off this card inside vmware esx 4.0. Using iometer I was able to generate 670 megabyte/sec read speed using a 100% sequential, 100% random, 4kb block size iometer test.

In fact, here are the other stats I came up with:

8x1.5TB Raid-0, 512kb stripe

**bonnie++ test**:
836mb/sec write,651mb/sec read,299cps, default block size

*IOMETER tests**:
a) 100% random, 100% read, 100 outstanding io, 4kb block
IOPS=167791
Tot.MB/sec=657
b) 100% sequential, 100% read, 100 outstanding io, 4kb block
IOPS=172442
Tot.MB/sec=673mb
c) 100% random, 100% read, 100 outstanding io, 512byte block
IOPS=175202
Tot.MB/sec=85.5mb/sec
d) 100% seqential, 100% read, 100 outstanding io, 1 mb block
Tot.MB/sec=5569mb/sec
IOPS=5569
e) 100% sequential, 100% write, 100 outstanding io, 4kb block
IOPS=107307
Tot.MB/sec=419mb/sec
f) 100% random, 100% write, 100 outstanding io, 4kb block
IOPS=1332
Tot.MB/sec=5.21mb/sec
g) 100% random, 100% write, 100 outstanding io, 512bytes block
IOPS=1663
Tot.MB/sec=0.81mb/sec
h) 100% random, 100% write, 100 outstanding io, 64kb block
IOPS=1481
Tot.MB/sec=92.6mb/sec
i) 100% random, 100% write, 100 outstanding io, 1mb block
IOPS=488
Tot.MB/sec=488mb/sec

[Computurd]

[Computurd]

nice link trans am, however i believe that was with eight HDD, results may be different, which also brings the question how much different will acards be vs ssd? but i think they will be WAY closer than acard to HDD

[trans am]

nice link trans am, however i believe that was with eight HDD, results may be different, which also brings the question how much different will acards be vs ssd? but i think they will be WAY closer than acard to HDD

GOD I feel dumb I didnt even notice the 8x1.5TB Raid-0, 512kb stripe. ha! I just assumed the poster was using ssds. I mean why even have one of these if you're drive is the bottleneck?? Seems like bringing sand to the beach to me. I am so excited to see steves acards in action. I have learned so much this week about storage in this thread I cant thank you guys enough. when it comes to advanced storage i've been in the clouds. I've always focused my attention on ram timings and overclocking cpus and gpus and all i've known for the most part it 2 raptors in raid0 with the best stripe size my controller could handle. then I got the ssds this summer and ive been frusterated because i could never fully see their potential using onboard raid.
I never even heard of dynamic ram disks until today when steve started the thread. I just want to see them in action. he's using 16 2GB dimms / 4GB in each ramdisk. is there a utility that shows the ram timings and speed of the ram in each disk? is the ram all the same speed and model or do you have mixed kits in there? this is all new to me. steve do you have any decent ssds you will be comparing to on the same test bed? It would be rad to see you compare to a popular common ssd like ocz vertex turbos. I wish they would loan you a few drives to compare with. These tests are a pain in the ass and very time consuming so as long as you're already doing it..... you know what I mean? I know your purpose in all of this is to compare SAS cards but..........

[NapalmV5]

and the fastest controller?? arrrrreeeca 1231!!

what i call the nostradamus effect

happy holidays

[Computurd]

lol i knew this thread would draw out the usual suspects lol....
really this will be a great chance to compare the cards and see what real world usage is. i guess the best thing to tell is going to be load times because we can say that 'x' solution loads 'x' game in so many seconds, seeing which is faster and by what percentage. if we go too far into 4k and 8k and 16k and 64k results then we will get into debates about which file size is more relevant, and under what usage pattern said file size is even relevant.
so there are too many points to argue in reality.
the only thing is that we can say what apps and games and things like that load and compare load times from stevero's system ONLY!!! and we will have to trust in steveros methodology that we can run these three in similar scenarios and setups giving us a true picture. I for one have total faith in ol steve-o!

[Tiltevros]

there are soo soo many things that we have to take in notice... soo many things my setup will be upgrade somewhere next week. we can test our controllers in our brand new 5520 chipset so stay tune.

[SteveRo]

Questions from Trans - is there a utility that shows the ram timings and speed of the ram in each disk? is the ram all the same speed and model or do you have mixed kits in there? this is all new to me. steve do you have any decent ssds you will be comparing to on the same test bed? It would be rad to see you compare to a popular common ssd like ocz vertex turbos. I wish they would loan you a few drives to compare with.

There is no utility that I am aware of that shows what the ram is doing inside the acards - that would be nice though. Based on some reading several months ago seems like I saw that the ram is all run 667 or slower and with lax timings, I suspect the reason for that is to keep compatability. I probably have at least 5 different make/models of DDR2 in this array. I only found one set that would not work and it wouldn't work in my 680i mobo either (bad memory). I do have some SLC mtron mobi 3000's - seq - 100/80, random - even slower - but no vertex

Question for everyone -

1. Please post iometer test profiles (.icf files) that you would like for me to use. Seq, random, varying xfer size, varying Q depth, workstation pattern, ... Much thanks to GullLars - use his reco for small files - http://dump.no/
2. Tell me or post a link that shows where/how to turn on the level load timer in crisis-whd.

[SteveRo]

For iometer I plan to run each controller -

1. Seq reads as a function of xfer size - from 4KB, 16, 32, 64, 128, 256, 512, 1MB, 2, 3, 4 for QDepth of 32
2. Seq write (same as above)
3. Random reads (same ...)
4. Random write (same ..)
5. 4KB random reads as a function of queue depth, qdepth=1, 2, 3, ... to 20 then 32, 64, 128, 512
6. Workstation pattern - 64KB 80% read, 20% write, 100% random at qdepth= 4, 16, 64, 256