>
> As an experiment I tried turning on Windows disk compression, the
> rational being smaller amounts of data back and forth from disk would
> result in relatively faster IO at the expense of CPU which I appeared
> ot have excessive amounts of. No such luck; my benchmark scratch
> compile (about 880 Java classes) went from 22 seconds to about 34
> seconds. By comparison, a single 2.8 Ghz CPU system with 10K RPM
> disks configured RAID 0 can do the compile in around 10 seconds.
>
As an aside, strictly sequential I/O with medium to large direct block
transfers is by far the fastest I/O. Anything that disturbs the
"sequential, large batch" nature causes massive degradations, which is
why even a DBMS needs large main memory buffer pools to work well.
