Write IOs are often returned to the host once the block is written in
the controller's NVRAM. Later, they are flushed to disk during the
destage process. However, read IOs may need to go to disk more
often. Similarly, two requests from a single stream may have widely
varying latencies if one hits in the cache and the other misses. In
certain RAID systems [5], writes may take four times
longer than reads due to parity reads and updates. In general, IOs
from a single stream may have widely-varying response times, affecting
the latency estimate. Fortunately, a moving average over a
sufficiently long period can absorb such variations and provide a more
consistent estimate.
Ajay Gulati
2009-01-14