On 2021/01/13 15:10, Martin K. Petersen wrote: > > James, > >> A co-operative device can help you find the optimal by returning >> QUEUE_FULL when it's fully occupied so we have a mechanism to track >> the queue full returns and change the depth interactively. > > Many disk drives and SSDs only return QF/TSF when they are out of actual > command slots (so typically 255). There is often a watertight barrier > between the transport-specific command processing and the media > management code. Very sad. > > So while the exponential backoff approach works great for disk arrays, > it's hit or miss with drives. And it's really hard to come up with a > good heuristic since the optimal queue depth is highly workload and > device-specific. > > We had a few folks looking at managing write queue depth based on > completion latency a couple of years ago but it didn't come to fruition. And if IO priorities are in use... And there is the latency target specs coming soon to make things even more interesting :) -- Damien Le Moal Western Digital Research
