Random Performance – Iometer (IOPS/latency), YAPT (random)

We are trying something different here. Folks tend to not like to click through pages and pages of benchmarks, so I'm going to weed out those that show little to no delta across different units (PCMark). I'm also going to group results performance trait tested. Here are the random access results:

Iometer:

Iometer is an I/O subsystem measurement and characterization tool for single and clustered systems. It was originally developed by the Intel Corporation and announced at the Intel Developers Forum (IDF) on February 17, 1998 – since then it got wide spread within the industry. Intel later discontinued work on Iometer and passed it onto the Open Source Development Lab (OSDL). In November 2001, code was dropped on SourceForge.net. Since the relaunch in February 2003, the project is driven by an international group of individuals who are continuesly improving, porting and extend the product.

Iometer – IOPS

iometer-ios-file.png

iometer-ios-web.png

iometer-ios-database.png

iometer-ios-workstation.png

When I first created these charts, I spent a few minutes making sure I had not entered the wrong data. This is just down right embarrassing for the rest of the field. Sure the P3700 did the same thing to other SSDs back when we tested it, but that was nearly a year ago. I figured nine months of newer PCIe SSD releases would at least close the gap a little bit, but I was clearly mistaken. Even at QD=1, the extremely low latencies result in nearly double the IOPS of all competing units. To put it simply, this means that any single request issued to the SSD 750 will complete *twice as fast* as anything else out there, and that's with competing PCIe SSDs included in the results!

To make matters worse, I should point out that the Iometer configuration used for these tests (unmodified here to retain equivalent results) pegs its worker thread at just over 200,000 IOPS. This means the SSD 750 would go even higher in at least three of these charts – if it wasn't so busy pegging a CPU core! In fairness, any single app that was applying heavy IO to an SSD would likely saturate its storage handling thread at a similar maximum IOPS. Moral of the story: If you want to peg an SSD 750, you had better be running a whole lot of, well, everything you have installed on your system. All at once.

Iometer – Average Transaction Time

For SSD reviews, HDD results are removed here as they throw the scale too far to tell any meaningful difference in the results. Queue depth has been reduced to 8 to further clarify the results (especially as typical consumer workloads rarely exceed QD=8). Some notes for interpreting results:

  • Times measured at QD=1 can double as a value of seek time (in HDD terms, that is).
  • A 'flatter' line means that drive will scale better and ramp up its IOPS when hit with multiple requests simultaneously, especially if that line falls lower than competing units.

iometer-trans-file.png

iometer-trans-web.png

iometer-trans-database.png

iometer-trans-workstation.png

I normally don't comment here, but just look at how much lower the IO latencies are for the SSD 750. NVMe absolutely has its perks, and Intel's 18-channel controller is certainly taking full advantage of it.

YAPT (random)

YAPT (yet another performance test) is a benchmark recommended by a pair of drive manufacturers and was incredibly difficult to locate as it hasn't been updated or used in quite some time.  That doesn't make it irrelevant by any means though, as the benchmark is quite useful.  It creates a test file of about 100 MB in size and runs both random and sequential read and write tests with it while changing the data I/O size in the process.  The misaligned nature of this test exposes the read-modify-write performance of SSDs and Advanced Format HDDs.

yapt-randreads.png

yapt-randwrites.png

This test has no regard for 4k alignment, and it brings many SSDs to their knees rather quickly. As we mentioned earlier, the SSD 750 is heavily optimized for 4k aligned writes, which explains the inconsistent results in this test.

« PreviousNext »