[ Thanks to Jeremy
Andrews for this link. ]
“With his recent release of 2.5.59-mm5, Andrew Morton explained
the anticipatory I/O scheduler that he has recently merged from
Nick Piggin. The resulting thread is an informative read,
explaining how and why this noticeably increases performance.“Much of the benefit gained from anticipatory I/O scheduling is
from the fact that read operations are usually synchronous, meaning
that the first read has to happen and report back before the next
read is scheduled. Without this recent patch, a command requiring
multiple synchronous reads gets each of them scheduled at the back
of the queue one at a time, resulting in noticeably poor
performance during streamed writes. With this patch, the reads are
moved toward the front of the queue, and a few millisecond pause is
actually added ‘anticipating’ the next read request. Andrew’s full
explanation follows, going into much more detail…”