Parallel tests add all tests to one partition if test durations are 0

[Gary Poster]

This is really only particularly important if tests are miscalculated to take 0 seconds, which they have in the past.

In that case, all tests can be put in a single partition.

AFAIK, this is the sort of thing I think we want, modulo prejudices against lambda.

=== modified file 'testrepository/testcommand.py'
--- testrepository/testcommand.py 2011-11-01 15:24:36 +0000
+++ testrepository/testcommand.py 2012-01-10 16:34:01 +0000
@@ -208,10 +208,12 @@
         # Scheduling is NP complete in general, so we avoid aiming for
         # perfection. A quick approximation that is sufficient for our general
         # needs:
- # sort the tests by time
+ # sort the tests by time, and then by len of the partition's tests.
         # allocate to partitions by putting each test in to the partition with
         # the current lowest time.
- queue = sorted(timed.items(), key=operator.itemgetter(1), reverse=True)
+ queue = sorted(timed.items(),
+ key=lambda item: (item[1], -len(item[0])),
+ reverse=True)
         for test_id, duration in queue:
             timed_partitions[0][0] = timed_partitions[0][0] + duration
             timed_partitions[0][1].append(test_id)

[Robert Collins (lifeless)]

For clarity, LP suffered from the 'tests get 0 duration' bug in an older testr; this led to a dual-core machine having a partition like:
[
  [0][testa, testb]
]
rather than
[
 [0][testa],
 [0][testb],
]
which would parallelise better in this corner case.

[Robert Collins (lifeless)]

Fixed in trunk. The sort that needed changing was the second one - the first sort sorts the tests to be in descending duration (so large chunks of time are allocated first), the second one in the inner loop selects the next partition to allocate to.