From: Godmar Back (gback@cs.utah.edu)
Date: Fri Dec 18 1998 - 20:06:31 EST
I checked in Jason's latest version. It now no longer assumes control
of the virtual address space, and should work well with libraries etc.
It does no longer use mmap, though the name is kind of a misnomer now.
(I guess that's then like gc-incremental.c, which isn't incremental.)
Note that I made it the default now. If we don't get adverse reports,
we will simply scrap the old hash scheme unless somebody comes up with
a reason why we shouldn't. Then we'll give Jason free hand in merging
gc-mem.c and gc-block-mmap.c, because the latter is technically part of
the former.
Here's some numbers I got for our unscientific benchmark compiling
the class libraries with pizza. These numbers were obtained by configuring
with the --with-timing option that Jason added. (Very neat, btw.)
"NEW" means with the new mmap scheme, and "OLD" means with the old hash
scheme. (You can ignore the "SYSTEM TIME" column).
NEW with -ms 8M -mx 64M:
FACILITY USER TIME SYSTEM TIME CALLS
--------------- --------------- --------------- ---------------
gc 5.496036 0.001698 5
TOTAL 23.193394 0.801650
NEW default (which is -ms 1M -mx 64M:)
FACILITY USER TIME SYSTEM TIME CALLS
--------------- --------------- --------------- ---------------
gc 26.550766 0.030335 36
TOTAL 45.376603 0.741702
OLD with -ms 8M -mx 64M:
FACILITY USER TIME SYSTEM TIME CALLS
--------------- --------------- --------------- ---------------
gc 6.386331 0.000636 5
TOTAL 24.204576 0.889243
OLD default (which is -ms 1M -mx 64M:)
FACILITY USER TIME SYSTEM TIME CALLS
--------------- --------------- --------------- ---------------
gc 34.323341 0.085792 40
TOTAL 53.319515 0.819459
In other words, if we provide sufficient memory to reduce the
number of garbage collections, there is only a speed up of 14% in
gc time, which amounts to an overall speed up of about 4.5%.
As we increase memory pressure and increase gc frequency,
the time spent in garbage collection is reduced by 23%, giving
an overall speedup of about 15%. I think this is a big
improvement.
Comments are welcome.
- Godmar
This archive was generated by hypermail 2b29 : Sat Sep 23 2000 - 19:57:23 EDT