|> I also am curious about the software overhead of a disk access, which
|> does not seem to be discussed very often, since seek time and rotational
|> latency tend to dominate disk transactions.
|>
The software overhead tends to be quite high these days, to the point
of being significant. If you spend around 2ms in the top end of a driver
getting a transfer set up for a device with a 12msec seek time and
a rotational latency of near zero (say, for SCSI disk with Write Cache
enable set), the software overhead would be significant (IMHO).
The software overhead was high enough for managing the the ESP (Emulex SCSI processor) and its 8-10 interrupts/SCSI command on the Sun systems I worked
on that it was more than worth the effort to try and cluster operations
together into a single read or write operation (in order to match the
