LSI MegaRAID 9271-8i PCIe Raid Controller Review

Sandy Bridge Testing – ATTO Disk Benchmark

The Sandy Bridge test system is PCIe 2.0 system. So obviously this is slower than the intended destination for this card. However it happens to be the machine I use every single day for the bulk of my work and testing. The testing quickly showed how optimized this card is for I/O performance. Maxing the PCIe bus to its limits. Prompting me to obtain a second system for comparison during this testing.

ATTIO Disk Benchmark

Baseline Comparison for Raid 0

Queue Depth 4

The baseline numbers are not bad for low queue depth they are on par with a single drive. The read numbers are exceptional they are about what you would expect the numbers to be sitting at if you added the advertised read and write numbers together.

Queue Depth 10

The greater queue depth Brings the numbers up a decent amount for the writes compared to to the first run. The Reads pretty much stay the same, they pick up a tiny bit.

So now that you have a comparison point lets move onto the drives on a real RAID controller.

Queue Depth 4

The small files benefit quite a bit from the controller here in the write department almost Doubling the performance. Its not till you get out of the small file sizes that you really see the extra read and write performance of the array come into play. Not to mention the additional write caching provided by the controllers on board cache. The Writes are nothing short of exceptional for this standard level of on drive cache usage.

Queue Depth 10

As one would expect with the larger queue depth. The smaller files benefit quite a bit from more being queued at once. With a much higher overall performance for the smaller files compared to the first run.

Queue Depth 4

Queue Depth 10

For those of you still following the data at this point You’ll notice that the results for the 5 drive stripe are pretty much identical to the two drive stripe. This is due to bus saturation by the controller. I spent a decent amount of time checking this before I was willing to settle on this conclusion. The stripe in this setup also has a high degree of internal diminishing returns between the drives due to the duplicate writes, however its not enough to negatively effect the performance and cause the numbers to be artificially low.

Spindled Drives

Queue Depth 4

To be expected the read performance is where the spindled drive really show a difference in performance. its about 1/4th the speed of the writes for the initial small file size runs. Unlike SSD’s the hard drives start making almost immediate use of the on drive cache. It does show for initial file size quite as much, but it shows well enough as we progress through the test. A lot of these numbers are due to the on controller cache. Though the combination of the high rpm and write speed work well for large files.

Queue Depth 10

We some some increase in performance with the larger queue depth. But not quite to the same multiples as you see on the SSD’s. With the fact that these are new writes and reads there is little if any benifit on either of these benchmarks from CacheCade. There just isn’t enough time for the system to optimize the heavy use and make use of the SSD caching. At least in default configuration as a read cache.

If i had a pool of three or so SSD’s in CacheCade configured for write back (read and write operations) it would probably benefit to a degree. However being as I only had commodity SSD’s on hand for testing. Its generally a bad idea to put them into use in these situations. As its an extremely easy way to kill an SSD quite quickly. They just aren’t rated for that degree of read and writes.

Summary of ATTO Disk Benchmark on Sandy Bridge i7 test setup

In this setup its really the test system showing its inability to cope with the extreme amount of data being thrown at the bus. Its really quite surprising. Its a great example for those currently with similar setups thinking about this card as a controller upgrade. Its amazing but its just way to much card for the hardware.