summaryrefslogtreecommitdiffstats
path: root/timings/JTimer
diff options
context:
space:
mode:
authorSteven Knight <knight@baldmt.com>2009-12-10 06:19:43 (GMT)
committerSteven Knight <knight@baldmt.com>2009-12-10 06:19:43 (GMT)
commit17879c1de6f589009c781e7f36e8a6a95b45c33f (patch)
treeab1e71c63d6a89ea409408287609180f6cf1acfe /timings/JTimer
parente8551494b473640e3310c8acf87bddaaf9804613 (diff)
downloadSCons-17879c1de6f589009c781e7f36e8a6a95b45c33f.zip
SCons-17879c1de6f589009c781e7f36e8a6a95b45c33f.tar.gz
SCons-17879c1de6f589009c781e7f36e8a6a95b45c33f.tar.bz2
Add a script for calibrating settings for timing configurations.
Update the timings scripts with calibrated settings that run a full build between 9.5 and 10.0 seconds on the buildbot slave.
Diffstat (limited to 'timings/JTimer')
-rw-r--r--timings/JTimer/TimeSCons-run.py29
1 files changed, 24 insertions, 5 deletions
diff --git a/timings/JTimer/TimeSCons-run.py b/timings/JTimer/TimeSCons-run.py
index 05ffbfb..7fe1bf4 100644
--- a/timings/JTimer/TimeSCons-run.py
+++ b/timings/JTimer/TimeSCons-run.py
@@ -25,10 +25,10 @@
This configuration is for timing how we evaluate long chains of
dependencies, specifically when -j is used.
-We set up a chain of 500 targets that get built from a Python function
-action with no source files (equivalent to "echo junk > $TARGET").
-Each target explicitly depends on the next target in turn, so the
-Taskmaster will do a deep walk of the dependency graph.
+We set up a chain of $TARGET_COUNT targets that get built from a
+Python function action with no source files (equivalent to "echo junk >
+$TARGET"). Each target explicitly depends on the next target in turn,
+so the Taskmaster will do a deep walk of the dependency graph.
This test case was contributed by Kevin Massey. Prior to revision 1468,
we had a serious O(N^2) problem in the Taskmaster when handling long
@@ -38,7 +38,26 @@ to the Taskmaster so it could be smarter about not re-evaluating Nodes.
import TestSCons
-test = TestSCons.TimeSCons(variables={'TARGET_COUNT':500})
+# Full-build time of just under 10 seconds on ubuntu-timings slave,
+# as determined by bin/calibrate.py on 9 December 2009:
+#
+# run 1: 3.211: TARGET_COUNT=50
+# run 2: 11.920: TARGET_COUNT=155
+# run 3: 9.182: TARGET_COUNT=130
+# run 4: 10.185: TARGET_COUNT=141
+# run 5: 9.945: TARGET_COUNT=138
+# run 6: 10.035: TARGET_COUNT=138
+# run 7: 9.898: TARGET_COUNT=137
+# run 8: 9.840: TARGET_COUNT=137
+# run 9: 10.054: TARGET_COUNT=137
+# run 10: 9.747: TARGET_COUNT=136
+# run 11: 9.778: TARGET_COUNT=136
+# run 12: 9.743: TARGET_COUNT=136
+#
+# The fact that this varies so much suggests that it's pretty
+# non-deterministic, which makes sense for a test involving -j.
+
+test = TestSCons.TimeSCons(variables={'TARGET_COUNT':136})
test.main()