the cost of soft updates, and the cost of hw.ata.wc=0
enclosed is a .jpeg of an xgraph of the following interactive test:
a monitor computer, rsh's to a test machine and does a 'date' command,
then waits 5 seconds, and repeats.
on the test machine, an athlon 1.2gig / 266fsb / 512meg/133(266ddr) /
wdc300ab udma100 drive, iwill mother board, an 'iozone 1024' is launched
to local disk (old iozone 2.0.1), just after the third rsh (ie, after
10seconds).
the y-axis of the chart is the time it takes the test machine to respond to
the 'rsh date'. the x-axis is the absolute time from the start of the
experiment(s). there are 3 sets of data:
hw.ata.wc=0, soft-updates enables.
hw.ata.wc=0, soft-updates disabled.
hw.ata.wc=1, soft-updates disabled.
the 'points' in the graph are the only real data (the lines
are xgraph's interpolations).
the data effectively shows that, under this (high disk) loading
condition, soft-updates are slower than no soft updates for a
(competing?) trivial interactive command. it also shows the HUGE cost
of turning write-caching off. think of the interactive pain as being
the area under the curves...
the small (blue) curve is what 4.1.1 'felt' like, the biggest (green)
curve is the (effectively) recommended mode of operation.
it seems to me that this says that freebsd's long-standing
characteristic of better performance than linux under load is weakened,
with the default configuration of 4.3-RELEASE.
i'll be pleased to provide the test (tcl) script, and my raw data, to anyone
that would care to see them, and i'm open to criticism about my
methodology. i also agree in advance that different stress conditions
would have changed the results at least somewhat.
-elh
<<attachment: interactive.jpg>>
