IOmeter testing

[Anvil]

Here's a small sample of the performance graphs I'm working on.

1 single Vertex 2 240GB
1 iodrive (One_Hertz)
1 LSI 9260-8i w/FastPath 4R0 Vertex 2/LE 100GB

I'll probably create a new thread with a different structure for easy comparison.

database_qd128_iodrive.png

a few more samples are coming up shortly...

fileserver_qd128_iodrive.png

webserver_qd128_iodrive.png

[One_Hertz]

Need some ACARDs in there to compete with the iodrive! Get SteveRO to run it on his beastly ACARD array. Are you using incompressible data in your testing? Vertex results look like they were done with zeros.

[Anvil]

Yes, the Vertex results (so far) have been using repeating bytes, pseudo random is coming up. (pseudo random is incompressible)

It doesn't make that much of a difference at qds 1-2, also keep in mind that the SF controller excells at mixed IO, random read is not that much to talk about vs e.g. the Intel X25 series.

I'll pm Steve for some action

The workstation graph includes a run with the V2 240GB using incompressible data.
Most real life data isn't incompressible, most databases are actually easily compressible.

Removed due to invalid results, test length to small on the Workstation tests, new one coming up...

[One_Hertz]

The red line looks very odd... It is impossible to see more than double the result from QD1 to QD2. Test size too small? Some of your profiles are missing a zero in the test size by the way so only 200MB is being used instead of 2GB.

[Anvil]

The size of the testfile has been 4x the size of the cache on the LSI 9260, so there is a small bit of cache in the results.

The red line in the lates graph is the WB cache on the LSI, it totally changes the way iops are handeled, WT is Write Through of course.

I'll check the profiles for the missing 0, the size of the testfile used is still correct (I've checked while doing the tests).

[One_Hertz]

The WB results are impossible. The results are saying that the average response is less than 0.05ms on QD2. The drives are simply not capable of this. There is an error somewhere in that test. (all the workstation profiles you linked me use 200MB as test size.)

[Anvil]

I've checked and the Workstation QD1-16 test was missing one 0 as you pointed out.

I'll redo the test on the 9260 and see what happens, still the size of the testfile was 2GB but that doesn't mean that the test was performed on the full size, I'll check.

[Anvil]

I didn't get to test it on the 9260 as that pc already has the 9265 installed, the 9265 is w/o FP key but I was quickly able to verify that the workstation result were invalid, at least for the 9260.

Will redo the Workstation tests as as I get to them, not a big issue, meanwhile the Database pattern is fairly close to the Workstation pattern.

edit:

ioDrive Random read 4K-128K QD1-128
rr_4K_128K_ioDrive.png

The ioDrive peaked 127K iops @ 4K QD128. (> 120K 4K iops @ QD32)

[Anvil]

A few more detailed graphs using QD1-16 linear.

Database_qd16_ioDrive.png

WebServer_qd16_ioDrive.png

Basically, there is no competition for the ioDrive outside raid controllers.

edit:

The LSI 9265-8i is tested with a testfile of 4x the size of the cache, 4R0 Vertex 2 100/120GB drives. (used drives)

There is no FastPath key on the LSI.
ioDrive_vs_LSI_9265_4R0_4GB_testfile.png

[One_Hertz]

I dont think Iodrive's low QD (1-4) performance can be touched by any SSD. ACARD is necessary. Iodrive also pulls far head in highly mixed small block read+write workloads(like fileserver profile).

[Anvil]

Could you correct the Workstation profiles and rerun those two tests?

[One_Hertz]

Could you correct the Workstation profiles and rerun those two tests?

It makes zero difference on iodrive whether it is 100MB or 80GB due to it having zero write caching. I didn't bother rerunning it in the first place due to this.

[Ao1]

Hi Anvil, any chance of including a 1MiB xfer in your test results

Whew, that iodrive is fast.

[Anvil]

Hi Ao1

You should try to make him an offer on that ioDrive , maybe there will be a bootable fusion-io drive someday.

The seqential transfer doesn't fit the iometer profiles due to the QDs.
One would have to create a new separate profile for sequential IO, using the most common sizes like 128KB and 1MB.

[One_Hertz]

I am NEVER giving up my iodrive. Hell, I am looking for a second one

[SteveRo]

Need some ACARDs in there to compete with the iodrive! Get SteveRO to run it on his beastly ACARD array. Are you using incompressible data in your testing? Vertex results look like they were done with zeros.

This next week sometime

[Computurd]

would be interesting to see a single acard drive, sans raid controller, btw mr steve-o. lets so those low qds 'unfiltered' on teh PCH! i bet theres some good stuff there

HOLY CRAP btw, number 36 in the world! LOL man you are awesome

[Anvil]

Really looking forward to some ACard action!

Look like he's been working hard on those points, we might see some more of him in the storage section when the new pcmv is realeased.
I wonder what impact the new version has on storage.

I'll be spending more time at hwbot when the X79 is released

[SteveRo]

Good morning all,

Iometer files located at -

https://docs.google.com/leaf?id=0B2A...thkey=CLer_rAK

One of the tests - I think - Random Read 4-128KB exp2 QD1-128_2011_aligned - locked up at test 6 for some reason, couldn't figure out why, all of the others seemed to run ok.

edit - all iometer runs with 2600k @ 45x

Some other tests just for comparison -

winsat with 2600k @ 45x



winsat with 2600k @ 50x



AS SSD - see screen shot for details of each -



Today or tomorrow I will start 1880/16x acard.
Much thanks to Mr. Anvil for pulling all this together!

[Anvil]

Hi Steve

What can I say, your results doesn't disappoint.

I've prepared the Database QD1-16 for now but the other ones will follow.

Strange that the random read test stopped, I'll look into where it failed.

Also, I've included a single Vertex 3, at low QD it performs as most SSDs, it does leave 2R0 Intel 320series in the dust while using compressible data.

Looks like One_Hertz could need that second ioDrive

database_qd16_sro_acard_pch.png

As we add more QD the graph changes slightly and the ioDrive is ahead of the ACards at QD128.

database_qd128_sro_acard_pch.png

[One_Hertz]

I am looking on ebay every day for a second iodrive, but I don't see anything

So at QD1 ACARDs are almost exactly as fast as iodrive? Looking forward to more graphs. For higher QDs I should be able to win if I find a second iodrive (or sell mine and get a 160gb iodrive duo).

[Computurd]

truly amazing results with those acards and the PCH! wow!

[Anvil]

These are great results.

Wonder whats going to be able to compete with ACards or the ioDrive.


fileserver_qd16_acard.png

fileserver_qd128_acard.png


WebServer makes it a bit more interesting, looks like the ioDrive is the winner as we are doing reads.

WebServer_qd16_acard.png

webserver_qd128_acard.png

[SteveRo]

I was planning on 3d benching a 580 under dice this weekend but I killed that card so I have been drowning my tears in iometer instead.

All tests below where with Areca 1880/4GB 16x Acard 9010 R0 with read ahead disabled, 2600k @ 45x multi

Read ahead disabled seems best for acards.
Based on AS SSD, any read ahead seems to just get in the way.
4k strip, 64K cluster (same as for the PCH).
Iometer test file size - complete drive filled, 28.4GB (0 GB free).
Edit - Turns out that these iometer runs where all at 2GB used, even though the file created was 28.4Gb - as such these runs are using the Areca 4GB DDR2 cache to unfair advantage
Iometer files located at -

https://docs.google.com/leaf?id=0B2A...uthkey=CIyYugY

The same test - Random Read 4-128KB exp2 QD1-128_2011_aligned - would not run beyond test 5 all the others ran to completion.

Data seemed more erratic but some of the numbers looked huge.

AS SSD and wei winsat disk for this config -

[SteveRo]

A few happy snaps -