🚀 go-pugleaf

[My apologies for the delay in reply; I got snowed under by real work.
 All of this data is for moderately used Ultrix 2.2 and 3.1 Vaxes.]
jps@wucs1.wustl.edu (James Sterbenz) writes:
[I mention surprising numbers from workload studies.]
| Sounds interesting...
| Can you give us some idea of what your surprising numbers look like?

 The buffer cache hit rate is very high (80% during a parallel kernel
remake on a client machine, approaching 92% on a lightly used
file/compute server). Even under heavy file I/O load (said kernel
remake) name-to-inode resolution accounted for a good 45% of the
block-level activity (more lightly-loaded systems see this soar to
50-60%), with cache hit rates of 96-99%. Asynchronous buffer writes
range from 58% of writes (during the kernel remake) to 25% (average
file/compute server activity). On the other hand, delayed writes are
very worthwhile; about 75% of them are never actually written to disk
during the kernel remake (presumably due to a second write filling the
buffer up). Even during a kernel remake on an NFS-mounted partition,
less than 21% of the blocks read or written were remote ones (at the
filesystem level, about 80% of the reads and 44% of the writes were
remote).

 There's an appallingly high level of symlink reading (probably due to
the filesystem shuffling everyone is doing to accommodate multiple
architectures and diskless clients); 19% of the system-level reads
during the kernel remake were symlink reads (almost all from
name-to-inode resolution), and this soars to 45% on our lightly loaded
central machine.

-- 
	"I shall clasp my hands together and bow to the corners of the world."
			Number Ten Ox, "Bridge of Birds"
cks@white.toronto.edu		   ...!{utgpu,utzoo,watmath}!utcsri!white!cks