Link Copied
Hello,
See for example this chart:
This chart represents utilization (as seen via util% in iostat(1)), in purple, and a performance metric, in green, measured during various benchmarks. (For simplicity we can assume the performance metric is proportional to the amount of IOPS)
As you can see, most of the time the purple and green line mostly overlap, which means, if util% is at 100% then performances cannot increase further. But in some instances, highlighted by the red circle, they do not overlap, which means that even if util% is at 100%, we can still get more throughput from the device.
The difference between them is that these within the red circle are single-threaded benchmarks, while those outside of the circles are multi-threaded.
My point is that "util%", which for HDDs is generally a good indicator of the device inability to provide more performance, is not useful for Optanes. From the chart we can see that a single-thread benchmark can push enough requests to be constantly waiting for the device response (pushing the util to 100%), but a second I/O thread can increase the throughput even further. Maybe also a third and fourth and so on.
This is not an issue with the performance, but rather an issue with the monitoring. This is probably common behavior for all SSDs that can serve multiple threads simultaneously (I think my specific model can handle between 2 and 4 threads with almost linear performance increase).
To summarize: given a system with a certain amount of workload, how do I find out how much more performance I can get? If starting more tasks will result in more throughput, or if I have reached the Optane performance limit?
