SM951 as OS boot device information

It looks like I found the answer to the low benchmark results.

Created a new RAID 0 of my SM951s, with a 16K stripe size. The result:




Does this look familiar?

One more thing to tweak. If you have the C States CPU power saving options Enabled in the UEFI, set them to Disabled before running the benchmark. IO is low CPU usage, so if the CPU is idling with C States enabled, that increases latency in IO.

PLEASE try using the 128K stripe size before giving up on these SSD and board. I'm hoping your mileage will NOT vary. Show me some great results... please!!

Ronzer, are you a happier SM951 RAID 0 user now? You're welcome...

Since I had to prove the 16K stripe was the culprit, I'll show you what a benchmark looks like with C States Enabled. Yes, I had C States Disabled for all my previous results.

Now, 16K Stripe with C States Enabled:



Hmm, lost over 10MB/s in 4K Read speed, and a bit everywhere else.

Intel previously had an option for IRST called DSA, Dynamic Storage Accelerator. That was an option in the IRST Windows program, and a UEFI option in Storage Configuration, to toggle On or Off.

DSA basically turned C States On and Off. They add latency to IO, even though IO is not CPU intensive. DSA would switch from C States on ("Low Gear" in Intel's dumbed down terminology) to C States off ("High Gear") automatically or manually by the user.

So where is DSA now? Good question. Removed in IRST 14 for some reason. Must be related to the PCIe SSDs. DSA is convenient for disabling C States, unless you use another tool for that.

Our ATTO results are similar in the shape of the Read speed results. You have better write results with double the NAND per SSD, and a Terabyte more of NAND to write to. Your Anvil results are great, I never ran that on my 128K stripe, better do that now.

So do you agree the 128K stripe is better than a 16K stripe? Imagine reading a 1 Gigabyte test file, 16K at a time from three SSDs. 128K is 8 times 16K. Plus underlying details of the NAND page size.

This is some of the first results in the wild with these SSDs in RAID 0. We know this stuff now, not many others do.

Hello! I requested BIOS update that:

• Allows booting with SM951 256Gb SSD through PCI-e adapter 

 for H87M Pro4 Motherboard from AsRock support.

Here is the BIOS they gave to me:

http://s000.tinyupload.com/index.php?file_id=02633572217212818984

It seemed to be custom BIOS dedicated for my request because it is not updated on site.
And I have no idea about BIOSes so don't know what they changed in it so, use it on your own responsibility.
All I know it works for me and I am running Windows 10 on my SM951.

TimH, I can't find which size SM951s you are using, but it seems to be three 256GB, is that right?

I tried a 64K stripe size to see if the larger the stripe size, the better the sequential/large file speed result trend would remain the same. With SATA SSDs, we don't see that trend as much since there is more available bandwidth, at least for up to three SSDs. That was true for Intel 8 and 9 series chipsets. The 10 series chipsets have greatly increased IO bandwidth, relative to SATA SSDs.

But the PCIe SSDs using an x4 interface seem to be using most if not all of the available bandwidth of the 10 series chipset (Z170) with two SM951s. I don't know if it is the IRST 14.x driver that is the limitation, or the bandwidth of the chipset. I read in an SM951 review on a very respectable PC review website that the real world maximum sequential speed of a PCIe 3.0 x4 interface will be ~3,200 MB/s. But IMO that does not seem right when using multiple x4 interfaces. I may be missing something here...

Regardless, my 64K stripe size result:



The trend seems to continue, the sequential read speed is slightly reduced from the 128K stripe size. The real world difference between 64K and 128K won't be particularly noticeable IMO. The 4K high queue depth speed seems somewhat better with the 64K stripe.

I forgot to mention that benchmark testing an OS volume will result in slightly reduced results, since it will be interrupted by OS tasks during the test.

It looks like the sweet spot for PCIe x4 SSDs in RAID 0 is a two drive array IMO. The increase in performance with a third SSD seems minimal. That is, at least for the Z170 chipset and the current state of the IRST RAID driver. The real world user experience with two and three drive RAID 0 arrays may be different, which only long term use of each would reveal. I'd like to see the benchmark results of a two, 512GB SM951 RAID 0 array.

Would you agree that the 16K stripe size is not optimal for use for SM951s in RAID 0?

128K and 64K are optimal for the best sequential performance, and don't sacrifice 4K performance very much, if at all. Those are the stripe sizes we would suggest to others, right?