In message <[EMAIL PROTECTED]> Ben Rosengart
writes:
: D'oh -- I *meant* to add "besides trying different values and measuring"
: -- if I had that much time on my hands, I wouldn't be worrying about how
: long a make world takes. :-)
Generally on FreeBSD machines that are otherwise unused and flush with
memory, the formula I've seen for n is 4 * #CPU. This appears to be
the sweet spot in the curves for people that have run experiments.
Beyond this point addtional jobs had no benefit or actually slowed
things down, at least according to the few tests that were run.
Of course there is the whole area file system tuning that also is
important (having enough spindles, using soft update, noatime, etc).
>From experiments I did in the 3.0 time frame, I found that file system
options could gain a PPro 200 about 10% for slow disks and as much as
25% for very fast disks.
The cc options also had a minor effect (on the order of 5% iirc).
I took an untuned make buildworld down from about 1:50:00 to 1:05:00
by using the above tuning (and building on newer, faster disks). My
machine wasn't flush enough with memory (only 96mb at the time) for me
to contemplate adding -j to the mix. Others have reported
improvements in the few percent range.
Warner
To Unsubscribe: send mail to [EMAIL PROTECTED]
with "unsubscribe freebsd-hackers" in the body of the message
