LSI MegaRAID 9271-8i PCIe Raid Controller Review

Sandy Bridge Testing – AS SSD

Continuing on from what LSI’s controller generally does the best of showing real world application in. We move on to Targeted SSD benchmarks.

I apologize at this point for repeating myself some however I am sure some of you skipping around will appreciate not having to go hunting for this information. Due to the way most SSD benchmarks work they don’t take advantage of the methods of drive control that a RAID controller generally uses. Since they are mostly looking for maximum read and write without heavy use of on drive cache for small files. A RAID controller tends to reflect poorly in these tests for that reason. However it is a standard method we use for testing storage on this site. So we are gonna to stick with the tried and true methods you are use to seeing.

AS SSD 

Baseline Comparison for Raid 0

As a starting point this is a good example of smile file transfer, with little or no queuing. The access times are fairly low and on par with some of the other SSD reviews posted on this site.

With a the baseline stripe size of 256KB we see that there is a decent amount of fluctuation in the benchmark results. The sequential read and write performance are much higher. The 4K performance falls off quite a bit on the write performance. The 4K-64thrd also shows decent improvement in this setup as well. The big difference is in the write access time. Overall the total score is much higher than on the Intel controller.

Stepping up the stripe size to 1MB we notice the numbers drop from the previous benchmark. While the 4k-64thrd goes up about 10MB/s. Access time stays mostly the same with a bit of natural fluctuation shown, nothing to show a difference there though. The overall score drops with the larger Stripe size.

Moving to a 16KB stripe size we see a drop in sequential performance, which is to be expected. We see dramatic improvement in 4K read performance, of around 10MB/s. The 4K-64thrd also shows significant increase in performance. Overall the score is closer to that of the default stripe size of 256KB.

In practical application you would see an increased performance in small file work loads with stripe size like this. Say for example if you where doing a large number of database writes.

Following in the foot steps of small stripe size being a preferred setup for this benchmark we move from the two drive setup into a five drive array. It shows close to triple the performance for sequential reads. the write performance doesn’t climb anywhere close to as much, at not quite double. The 4K read performance achieves a minuscule increase of around 4MB/s. The Write performance for 4K stays relatively the same enough that it can be considered to be the same as the previous run. The 4K-64thrd shows around 140MB/s increase in read performance, write performance again stays the same though. The access time once again remains constant.

Overall this is a good result of the controller scaling as drives are added to an array. The file sizes are small enough that we are not encountering diminishing returns from the added traffic created by the small stripe size.

For the spindled drives the numbers are not horrible. Really what you are seeing here is the controller doing its part to work as a cache and speed up the transaction process. While compared to an SSD these numbers are horrible in most situations. The reality that Read and write performance for sequential writes is a nice sign. The 4K and 4K-64thrd also show what one would expect to see from a hard drive. The numbers are somewhat higher than they would be without the raid controller. This again speaks well for the Raid controller.

Summary Of AS SSD Benchmark On Sandy Bridge i7 Test Setup

A SSD benchmark like this really access as showing small file performance. While in most situations you will be setting up the your stripe sizes around the software implemented. It does a good job of showing the flexibility of the controller to handle varying requirements.

In practice the additional level of software in a server would probably step in and improve these numbers quite a bit. However as a raw example they are a good baseline of performance gains that can be achieved.