Restore old initial batch distribution logic in LoadScheduling (#812)

* Restore old initial batch distribution logic in LoadScheduling

pytest orders tests for optimal sequential execution - i. e. avoiding
unnecessary setup and teardown of fixtures. So executing tests in consecutive
chunks is important for optimal performance.

Commit 09d79ace35 optimized test distribution for
the corner case, when the number of tests is less than 2 * number of nodes.
At the same time, it made initial test distribution worse for all other cases.
If some tests use some fixture, and these tests fit into the initial batch,
the fixture will be created min(n_tests, n_workers) times, no matter how many
other tests there are. With the old algorithm (before
09d79ace35), if there are enough tests not using
the fixture, the fixture was created only once.

So restore the old behavior for typical cases where the number of tests is
much greater than the number of workers (or, strictly speaking, when there
are at least 2 tests for every node).

In my test suite, where fixtures create Docker containers, this change reduces
total run time by 10-15%.

This is a partial revert of commit 09d79ace35

Co-authored-by: Bruno Oliveira <nicoddemus@gmail.com>

changelog/812.feature

@@ -0,0 +1,22 @@
+Partially restore old initial batch distribution algorithm in ``LoadScheduling``.
+
+pytest orders tests for optimal sequential execution - i. e. avoiding
+unnecessary setup and teardown of fixtures. So executing tests in consecutive
+chunks is important for optimal performance.
+
+In v1.14, initial test distribution in ``LoadScheduling`` was changed to
+round-robin, optimized for the corner case, when the number of tests is less
+than ``2 * number of nodes``. At the same time, it became worse for all other
+cases.
+
+For example: if some tests use some "heavy" fixture, and these tests fit into
+the initial batch, with round-robin distribution the fixture will be created
+``min(n_tests, n_workers)`` times, no matter how many other tests there are.
+
+With the old algorithm (before v1.14), if there are enough tests not using
+the fixture, the fixture was created only once.
+
+So restore the old behavior for typical cases where the number of tests is
+much greater than the number of workers (or, strictly speaking, when there
+are at least 2 tests for every node).
+

src/xdist/scheduler/load.py

@@ -248,13 +248,27 @@ class LoadScheduling:
         # Send a batch of tests to run. If we don't have at least two
         # tests per node, we have to send them all so that we can send
         # shutdown signals and get all nodes working.
-        initial_batch = max(len(self.pending) // 4, 2 * len(self.nodes))
-
-        # distribute tests round-robin up to the batch size
-        # (or until we run out)
+        if len(self.pending) < 2 * len(self.nodes):
+            # Distribute tests round-robin. Try to load all nodes if there are
+            # enough tests. The other branch tries sends at least 2 tests
+            # to each node - which is suboptimal when you have less than
+            # 2 * len(nodes) tests.
             nodes = cycle(self.nodes)
-        for i in range(initial_batch):
+            for i in range(len(self.pending)):
                 self._send_tests(next(nodes), 1)
+        else:
+            # Send batches of consecutive tests. By default, pytest sorts tests
+            # in order for optimal single-threaded execution, minimizing the
+            # number of necessary fixture setup/teardown. Try to keep that
+            # optimal order for every worker.
+
+            # how many items per node do we have about?
+            items_per_node = len(self.collection) // len(self.node2pending)
+            # take a fraction of tests for initial distribution
+            node_chunksize = max(items_per_node // 4, 2)
+            # and initialize each node with a chunk of tests
+            for node in self.nodes:
+                self._send_tests(node, node_chunksize)
 
         if not self.pending:
             # initial distribution sent all tests, start node shutdown

testing/test_dsession.py

@@ -115,18 +115,18 @@ class TestLoadScheduling:
         # assert not sched.tests_finished
         sent1 = node1.sent
         sent2 = node2.sent
-        assert sent1 == [0, 2]
-        assert sent2 == [1, 3]
+        assert sent1 == [0, 1]
+        assert sent2 == [2, 3]
         assert sched.pending == [4, 5]
         assert sched.node2pending[node1] == sent1
         assert sched.node2pending[node2] == sent2
         assert len(sched.pending) == 2
         sched.mark_test_complete(node1, 0)
-        assert node1.sent == [0, 2, 4]
+        assert node1.sent == [0, 1, 4]
         assert sched.pending == [5]
-        assert node2.sent == [1, 3]
-        sched.mark_test_complete(node1, 2)
-        assert node1.sent == [0, 2, 4, 5]
+        assert node2.sent == [2, 3]
+        sched.mark_test_complete(node1, 1)
+        assert node1.sent == [0, 1, 4, 5]
         assert not sched.pending
 
     def test_schedule_fewer_tests_than_nodes(self, pytester: pytest.Pytester) -> None: