Provide guidance for setting readers and writers

Thanks .

So here is the guidance for the documentation (for the Fixed Value bullet point):

"A good rule of thumb is set each of readers and writers equal to the queue depth of the underlying IO subsystem (i.e. readers = queue_depth and writers = queue_depth). However, for best performance it is recommended to benchmark with different settings and pick the one that best meets your throughput and latency requirements in your environment"

Dave Rigby one clarification needed is whether the (readers + writers) = queue depth or is it (readers = queue depth) and (writers = queue depth)?

As a basic rule for the documentation I would suggest the latter - for a couple of reasons:

1. Both readers (bgFetch, DCP backfill) and writers (flushing, compaction) are bursty - as such you probably want utilise the available queue depth for each task when it's running. If we only gave readers half the queue depth then when not flushing / compacting you'd likely not make use of the available disk bandwidth.

2. Writes arn't actually blocking (until we issue the final fsync in a batch) - as such you're not necessarily consuming any disk IO (and queue depth) all the time.

Ultimately, for optimal performance in a given workload / environment the user would need to experiment and benchmark the different settings and pick the one which meets their requirements (typically highest OP/s while still within their latency SLAs). I think a reasonable starting point is reader=writers=queue_depth.

 one clarification needed is whether the (readers + writers) = queue depth or is it (readers = queue depth) and (writers = queue depth)?

.  FYI