I wanted to mention that back around 1990 or so I spent a bunch of
time trying improve performance by playing with the disk sort routine
(the elevator alg) in SunOS. I futzed around for several weeks before
realizing that it didn't much matter what I did. That was the insight.
The reason is that UFS, at least, did a fantastic job making sure that
all the data you cared about was close together. So there was little
or nothing that you could do for the single user work station case,
the file system had, in effect, already sorted the requests before they
were sent to the disk.
On the other hand, it is certainly true that there is performance to
had or lost when the system is under heavy load. But even then I'd
strongly urge people that want to work on this to instrument the kernel
and gather some traces from a busy web server (or whatever it is you are
trying to make go fast) and stare at them for a long time, long enough
to understand why you are seeing what you are seeing. The reason is
that the file system's allocation policy, read ahead policy, meta data
placement, etc., all skew the data.
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/