Statistical puzzle

[Garrett Wollman]

<<On 16 Feb 2005 23:10:08 +0100, Tomas Olsson <tol at stacken.kth.se> said:

> Interesting. Could you share some more details, exactly how did you
> measure what, how did you exclude close/write-back times? Code?

<http://cvsweb.freebsd.org/src/bin/dd/>

Note that dd never calls close(); it simply writes continuously and
then exits.  (The statistics are printed by summary(), which is called
from exit() by way of atexit().)  So in the case of:

$ time dd if=/dev/zero of=/afs/foo/bar/baz bs=1024k count=1000
1000+0 records in
1000+0 records out
1048576000 bytes transferred in 24.841135 secs (42211276 bytes/sec)

real    2m46.852s
user    0m0.000s
sys     0m4.966s

We know that the first time is printed before the close (which only
happens implcitly as a part of exit), and the second time is printed
after exit (and therefore close) completes.  The difference in the two
times should be entirely due to the writeback, unless Arla is starting
the writeback before the close (which would be a good idea if a good
heuristic for sequential writes can be developed).

Unfortunately the Lignux dd command does not have this feature so it
is not possible to compare the two systems on this basis.

The statistical displays in my message were generated by Poul-Henning
Kamp's ministat programing, from src/tools/tools/ministat.

-GAWollman