Have you checked to make sure that the clustering is really working
in the 1K cases? Larry McVoy told me some time ago when I talked to
him about this subject that he had added a system call to use to
check to see if any clustering was happening, and he didn't see a
single instance of it actually happening.
Hmmm. I must admit that I just assumed that the clustering patches were doing
their job as they've been in the kernel for quite a long time now. I'll have
to read over the clustering code enough to understand how it works and run some
further tests.
I seem to recall that benchmarks made in the early days of the clustering
patches made against 4KB blocks that were virtually indistinguishable
from those made with 1KB blocks. Since this doesn't agree with your
benchmarks, it seems that before thinking too hard about adding fragments,
it would be worth seeing if the buffer cache needs any tuning to take
better advantage of clustering.
Agreed. I would much prefer to stay with 1K blocks if the same or nearly the
same performance can be achieved with them. However, I vaguely recall that the
clustering code wasn't supposed to matter that much if the SCSI driver
supported reasonable scatter/gather operations, and that's certainly the case
with the tests I reported on.
Leonard