Steady State PerformanceIn this test, we put the drive through a steady-state measurement of its write performance in which we write onto the drive with a non-stop, fully-randomized stream of data for an hour (in which the controller can't find patterns).
The drive's 4K write performance is measured over three levels of loading (amount of used space vs. amount of free space for the controller to use as a potentially spare performance-enhancing area). The three levels we used are 50%, 75%, and 100% LBA. The test may not be a suitable representation of a consumer scenario in which data is typically written in short bursts with lower QD, but can be of academic interest by showing how the various SSD controller architectures manage their NAND flash resources.
SSD controllers work in mysterious ways. Some also use free NAND flash area as performance-enhancing "scratchpads." Some modern controllers, such as the SandForce SF-2281 and LAMD LM87800, have a fixed amount of physical NAND flash known as the overprovisioned area allotted to these "scratchpads". The controllers can even use free space in the user area (the part of the SSD's NAND flash that isn't overprovisioned). As we begin our test, the SSD controller has vast amounts of free space/OP area at its disposal, so it pumps its highest write IOPS out then. Once the overprovisioned (OP) area is full, we see write IOPS drop like a rock until it plateaus, with the odd spike as the controller finds free space to use. Towards the end (the last 15 minutes), the SSD controller will have run out of such free space, providing a metric we call "steady state write performance". That phase of our test is of particular importance, and the second graph below focuses on that portion.
|Capacity Used: Compare with:|