Skip to content
Home > Blog > March Madness – do your SSDs perform like champions under pressure or are they “one and dones”?

March Madness – do your SSDs perform like champions under pressure or are they “one and dones”?

  • JB Baker 
  • 3 min read

With the NCAA March Madness tournament upon us, many will be cheering on their Cinderella teams and cursing the bracket-busting upsets of their highly seeded picks expected to win it all.  It will all come down to how well they perform under pressure.

Well, I asked my applications engineering team (thank you Keith McKay ) a similar question: How well do the ScaleFlux SSDs perform under pressure vs the big names? Turns out, we do really well both during the lighter pressure of the early rounds all the way through the intense pressure of the finals.

Okay, turning down the volume on the metaphor, here’s the scoop.

Test set up:

  • We compared the 7.68TB enterprise NVMe SSDs from ScaleFlux and a well-known vendor.  After preconditioning, we ran a 50/50 RR/RW workload against each drive.
  • Instead of just running for “max IOPs,” we ran pre-set pressure levels, with the workload generator attempting to complete 100k, 150k, 200k, etc up to 1100k IOPs.
  • In each case, FIO maintained a 50/50 ratio of reads and writes.  We recorded how many IOPs each drive achieved and what the 99% Read Latency was at each pressure level.

Results:

IOPs

  • The other drive saturated at the 350k IOPs pressure level.  Attempting to push more IOPs from the host would not produce any higher output from the drive.
  • The ScaleFlux CSD saturated at 950k IOPs -> 2.7x the max performance!

99% Read Latency

  • This gets even more interesting and a bit more complex. Refer to the graph below to see what I mean.
  • The other drive started off at 522us, climbing steadily to 634us as the pressure mounted to the IOPs saturation point of 350k IOPs.
  • The ScaleFlux CSD started off at a much lower level of 248us, gradually climbing to 987us at the 950k IOPs point.

Key data points from the graph

  • At every pressure level up to 350k IOPs, the other drive’s latency was 1.5-2x higher (that is, worse) than the ScaleFlux drive.
  • The ScaleFlux CSD’s latency was better than the other drive all the way to 750k IOPs.  That is over 2x the IOPs with the same or better latency!!
  • Only after 750k IOPs was the ScaleFlux CSD latency higher than the other drive.

What’s all this mean?

The ScaleFlux drive shines regardless of the pressure level, with 1.5x better latency a lower pressures, 2-4x better IOPs at any given latency level, and 2.7x the maximum IOPs. So, for your latency sensitive applications (I’m looking at you #AI, #machinelearning, #highfrequencytrading, #HTAP, #OLTP, #OLAP, #databases, #frauddetection), #CSD3000 from www.ScaleFlux.com is worth a look regardless of whether your at the low intensity of warm ups or the extreme pressure of the championship game.

#1 seed potential all the way!!

Solid lines show the IOPs Achieved

For further reading – discussion of the benefits of SSDs and impact of QoS and Latency on databases: Are SSDs good for databases? – Darwin’s Data (darwinsdata.com)

JB Baker

JB Baker

JB Baker is a successful technology business leader with a 25+ year track record of driving top and bottom line growth through new products for enterprise and data center storage. He joined ScaleFlux in 2018 to lead Product Planning & Marketing as we expand the capabilities of Computational Storage and its adoption in the marketplace.