Hello,
I have created a patch to improve make's behaviour with regards to
parallel execution and system load.
On Linux do the load averages refresh only every 5 seconds. Further do
load averages only ever represent past but not present load. As a
consequence does this lead to idle times during parallel builds where
make is waiting for the load averages to come down. It can also spawn
excessive amounts of processes (although it tries to balance it with
its own heuristic).
A better way to achieve a stable load is to use the active number of
running processes & threads over of the load averages. This number can
be found in /proc/loadavg on Linux and is updated in real-time. An
example:
$ cat /proc/loadavg
0.00 2.72 6.89 1/404 1890
Here it shows 1 (out of 404) running processes. The patch allows make
to use the 4th value from /proc/loadavg.
The overall improvement for a parallel build on an 8-core CPU:
-j -l 24 (old): 269.993128122 seconds time elapsed
-j -l 24 (new): 231.300406406 seconds time elapsed
In comparison:
-j 24: 230.482726977 seconds time elapsed
The behaviour of the -l <value> option with the patch is now much
closer to that of the -j <value> option and shows almost identical
build times for an empty system.
The test suite passes successfully with the patch.
You'll find two patch files against make-4.2.1 in the attachment. One
for configure.ac (includes a test for /proc/loadavg usability) and the
other for job.c (modifies the function load_too_high()).
Regards,
Sven
_______________________________________________
Bug-make mailing list
address@hidden
https://lists.gnu.org/mailman/listinfo/bug-make
proc-loadavg.patch