Subject: performance of "make build" in -currrent vs. 1.5
To: None <>
From: Chuck Silvers <>
List: tech-kern
Date: 02/09/2001 00:56:11
hi folks,

there were some reports a while back that "make build" was taking
substantially longer since UBC was integrated then it did with 1.5.
there have been a number of performance fixes to the UVM/UBC code recently,
so I just did the experiment myself, running "make build" of a 1.5-branch
source tree with both a 1.5-branch kernel and a -current kernel from
yesterday (plus the usage-balancing code I posted here yesterday).
the build didn't complete in either case because of errors in the
source tree being compiled, but it got far enough for the results
to be interesting:

3342.747u 2942.356s 2:27:02.46 71.2%    0+0k 51276+205546io 11027pf+0w

3428.843u 3304.275s 2:23:32.31 78.1%    0+0k 7645+36433io 2291pf+0w

so -current+balancing took 2.4% less clock time while using 2.5% more
user cpu time and 12.3% more system cpu time.  there's room for
improvement in the cpu usage here, but overall I would call this
a slight improvement over 1.5.  I'll be looking into reducing cpu
consumption in the near future (and fix the rusage numbers for reads
and writes, which are obviously not being updated correctly.)

does anyone else have any numbers to share about performance differences
between 1.5 and -current?