While synthetic 100% read or 100% write workloads do a great job of testing the underlying technology and reporting easy to understand results, they aren’t always indicative of how the drive will be used by the end user. Workloads that simulate enterprise environments try to bridge that gap without being overly complex. The process of measuring our server workload performance is the same as measuring random. The drive is first secure erased to get it in a clean state. Next, the drive is filled by sequentially writing to the RAW NAND capacity twice. We then precondition the drive with respective server workload at QD256 until the drive is in a steady state. Finally, we cycle through QD1-256 for 5 minutes each measuring performance. All this is scripted to run with no breaks in between. The last hour of RAID 0 preconditioning, the average IOPS, and average latency for each QD is graphed below.
The Database profile is 8K transfers, and 67% percent of operations are reads.
In RAID 0, during the Database benchmark latency stayed below 1.75ms on average and peaked out at 155K IOPS. If you look back to our review of the Intel DC P3700, while it isn’t the same system and our methodology changed a bit since then, during the database benchmark that it has a bit of an advantage at lower QDs until QD64. So depending on your specific application QD workload, the DC P3700 may be a bit of a better option from Intel in terms of performance, let alone write endurance.
The Email Server profile is similar to the Database profile, only it 8K transfers at 50% reads and 50% writes.
During the Email Server benchmark high QD IOPS averages were the highest we have seen from a single drive SSD. Latency was pretty well controlled until QD 128, after that it jumps from about 1ms to just over 2ms. IOPS consistency and latency was very good at QD256 as well. IOPS ranged around 10K and latency ranged 0.1ms.