LSI 9260 vs. AMD SB850

[Anvil]

1GB/s is easily achieved using 6Gb/s SSDs, F@32 was using 3Gb/s SSDs.

[F@32]

Those numbers on SB850 look a bit low. I don't know if it's the drives not playing nice, or some other factor. But I generally see >1GB/sec reads from SB850. http://www.xtremesystems.org/forums/...6&postcount=93
--Matt

Lol, my sequential are low, your randoms are low... Have you sorted that out yet? And yeah, I used 5R0 SATA2 drives which in theory should have pushed over 1Gb/s sequential read (3*200+2*250).

[Computurd]

LSI 9211 would only use ONE core on my C2Q build, and it was obvious it is exactly CPU limited. (it gave awesome numbers though :P)

i wonder if you could isolate that process and/or driver and force affinity and the priority via the task manager? would be interesting to try....


cant wait to see your results with the new card

[alfaunits]

i wonder if you could isolate that process and/or driver and force affinity and the priority via the task manager? would be interesting to try....

It is probably not possible the way the driver is written. If it has a global lock of some sort, it would be impossible to force multiple threads. (i.e. it could be doing the processing in its own system thread, a single one - nothing one can do here).

[F@32]

cant wait to see your results with the new card

Looks like MSM and drivers can be loaded from LSI. Not sure I will want to risk FW as well right away If it can match SB850 4K IOPS and show only 10% on CPUs - I will have desktop winner for 2R0/4R0.

[Ao1]

So without the fastpath key a single SSD can outperform an array when it comes to 4K random reads at QD1.

I think it really sucks when hardware is deliberately gimped with software.

It’s kind of strange that ICH performs so well with SSD as it predates mainstream SSD technology and could not therefore have been specifically designed to perform so well with SSD in mind. I bet Intel are kicking themselves.

[mattkosem]

Lol, my sequential are low, your randoms are low... Have you sorted that out yet? And yeah, I used 5R0 SATA2 drives which in theory should have pushed over 1Gb/s sequential read (3*200+2*250).

More or less, yes. The latest bios and drivers picked up the pep on pretty much all of the randoms, but at the cost of deep queue random read perf. The seq reads vary by 100MB/sec or so from test to test, so I wouldn't read too far into it.

--Matt

[F@32]

More or less, yes. The latest bios and drivers picked up the pep on pretty much all of the randoms, but at the cost of deep queue random read perf. The seq reads vary by 100MB/sec or so from test to test, so I wouldn't read too far into it.

--Matt

Nice, those are some good numbers. So, it looks like SB850 is capped at 1Gb/s or so, right?

[mattkosem]

Nice, those are some good numbers. So, it looks like SB850 is capped at 1Gb/s or so, right?

Yeah, roughly. The highest I've seen in any benchmark was 1149MB/sec on winsat. Others seem to top out around 1050 or so though. It looks like 1037MB/sec is the highest I've seen on it with CDM or AS SSD.

--Matt

[F@32]

So, I'm not alone to experience sub-par 9260 performance without FastPath...

The thread: 3R0 Vertex LE

And the results from the thread:

[Computurd]

run without always read ahead, and turn disk cache policy to direct. that will fix it somewhat at least thats where the guy linked went wrong.
what settings are you using?

[felix_w]

F@32 , i noticed that your comparisons involve 1000MB for LSI and 100MB for SB850 when running CDM.

Could that be a possible cause for inconsistency in performance comparison ?

I also have this LSI card and a pile of SB600, SB710, SB750 mainboards if i could help with any testing, i'll be glad to do it....

...though only tested with my SB750 ASUS, so i don't know if it really is compatible with the others.

EDIT : Here is my R-0 with 4x Vertex 30GB (FW1.5) - NoRA-WBC-DirectIO on LSI (fw i think is the first released one)
12h Bios Screen

[F@32]

run without always read ahead, and turn disk cache policy to direct. that will fix it somewhat at least thats where the guy linked went wrong.
what settings are you using?

That's his results, and from what I read in that thread, once he got FP and set everything as you recommended, it was all nice and smooth.

I returned my 9260 since without FP I didn't think its 2R0 performance justifies price premium over SB850 or PCH no matter which settings I used. Of course what ASSSD and CDM don't tell you is that 4K QD32 will tax CPU at over 30%, while 9260 keeps it under 10%.

F@32 , i noticed that your comparisons involve 1000MB for LSI and 100MB for SB850 when running CDM.

Could that be a possible cause for inconsistency in performance comparison ?

Possible, but not very likely. Drives were HDDErased before each HW change. And 9260 needs 1Gb test run, otherwise I'd have been benching cache (512MB on-board). The results were not totally surprising, since it's hard to beat ICH or PCH with 2R0. LSI is meant for massive throughput with 4+ devices. For OS/apps purposes I plan to stay on 2R0, and I was looking if there was anything faster while not costing a grand.

See below (I believe all were done with 64K strip):

[Computurd]

F@32 , i noticed that your comparisons involve 1000MB for LSI and 100MB for SB850 when running CDM

Lol, how in the world did i miss this???

jesus, that makes a WORLD of difference!

[felix_w]

Lol, how in the world did i miss this???

jesus, that makes a WORLD of difference!

...been doin' a lot of practice lately in the "Find the differences" game....

[F@32]

Lol, how in the world did i miss this???

jesus, that makes a WORLD of difference!

I see that 9260 with 100Mb test would show close to 1.5Gb/s cache performance. But how does it impact cacheless runs? I will run P67 with 1Gb and 100Mb to see if it matters.

[felix_w]

Another thing we could see is the performance of SB850 @ CDM for 1000MB test, maybe it shows interesting things.

If you have time and the setup is ready, it would be great.

[F@32]

EDIT: I noticed that LSI chose to gimp 4K performance of cards without FastPath. Looking at Felix's 9260, that chap on OCZ forums and mine 9260/9240 it is very evident. Now, we know that FastPath unleashes the beast, but vanilla 9260 not being able to beat SB850 in QD1-32 4K writes? It's a joke. So, not gimped 9260 is actually $500 part, not $350...

Another thing we could see is the performance of SB850 @ CDM for 1000MB test, maybe it shows interesting things.

If you have time and the setup is ready, it would be great.

Felix, sorry my SB850 was sold. I only have P67 right now and LSI 9240 laying around. Anyway, here's back to back CDM on P67 with 100Mb and 1Gb test runs. Note that this is dirty drive with OS installed even though test was performed while I booted from another partition. As I suspected it doesn't matter much for cacheless runs. I can show AS SSD scores they also point that on-board RAID is faster. But AS SSD shows really low LSI sequential reads, so I don't know if it can be trusted there...



And that's of course how 9260 looks like when testing cache vs. drives