From: Godmar Back (gback@marker.cs.utah.edu)
Date: Tue Jan 26 1999 - 02:31:33 EST
With the elimination of the Kaffe_ThreadInterface, my favorite benchmark
(compiling the kaffe library sources) is now down to ca. 16.0 user seconds
from 17.1 seconds. (A few months ago it was still 26 seconds, btw.)
Note also that I moved the signal stuff down in systems/*/signal.c
This is necessary since we'll do some signal magic to speed up
context switches. Actually, it's not so much magic --- it's just avoiding
setjmps that save the signal mask every time (like FreeBSD does).
Pat has found that doing so reduces the thread switch time from 1400
to 220 cycles. This applies to operations such as Thread.wait, MONITOR_ENTER,
and Thread.yield().
- Godmar
This archive was generated by hypermail 2b29 : Sat Sep 23 2000 - 19:57:48 EDT