Latency Weighted Percentile – Comparisons

This is the same data from the previous page, but this time, we are comparing all tested drives at a specified Queue Depth on each chart. I'm going to minimize writing between them so it is easier to scroll through and see the progression as QD ramps up.

Reads

When it comes to low latency reads, the Samsung 950 Pro (orange) cleans house. It maintains a ~12ns lead on both the RD400 (cyan) and SSD 750 (yellow) at every queue depth. Bonus points to the RD400 for being fast enough to battle it out with the SSD 750.

Had to zoom out here to properly show the Kingston's extra long tail latency here.

Writes

Compare the first chart below to the one above (which is the same scale). Reads require look-ups and retrieval from flash before the IO can be completed, but writes can be acknowledged as soon as the data is received by the SSD (it figures out where to put it later). The added steps needed for read completion cause the near 10x increase in read latency when compared to writes.

Apart from morphing into what resembles a NYC subway map, we can note how the RD400 is more inconsistent than most other SSDs tested. We want that corner high and to the left, and at higher queue depths, the RD400 takes a sharp turn at 63%, lower than most other tested SSDs. You can also observe how the RD400 (cyan) battles back and forth with the 950 Pro (orange), while the Intel SSD 750 (yellow) gobbles both of them up whole.

Earlier I promised to show the old method. Here is that last chart repeated without weighting:

Those almost don't look like the same results, do they? The M6e (blue) appears to have an issue, but not one bad enough to warrant the very low IOPS seen, and the 63% 'latency shelf' of the RD400 has completely vanished! It takes looking back at the old way to realize how IO-weighted results artificially enhance the results, making them look far better than they actually are.

« PreviousNext »