Visible benchmark regression between 08/05-09/05

[hubertp]

Most visible in

But similar bump is seen in other stdlib benchmarks.

[GregoryTravis]

I reproduced a large benchmark regression in this test (org_enso_benchmarks_generated_Collections_vector_float_fold). This unusual test uses List and does all its operations through Meta:

sum_list_meta list =
    nil_cons = Meta.meta List.Nil . constructor
    folder acc list =
        meta_list = Meta.meta list
        if meta_list.constructor == nil_cons then acc else
            fs = meta_list.fields
            @Tail_Call folder (acc + fs.at 0) (fs.at 1)
    res = folder 0 list
    res

Between f1ddf1b and 720d32c, this test got 25x slower, and this is consistently reproducible. The smallest change to reproduce this is shown below -- a change to use the new build instead of new_builder, in both Bench.build and Bench_Builder.group.

This is strange because it seems like it should only affect the construction of the benchmark groups, not the benchmarks themselves. Also strange is the fact that reverting both of the changes below restores the old speed, while reverting only one of these does not change the speed at all.

diff --git a/distribution/lib/Standard/Test/0.0.0-dev/src/Bench.enso b/distribution/lib/Standard/Test/0.0.0-dev/src/Bench.enso
index 42f24a8701..08994596f6 100644
--- a/distribution/lib/Standard/Test/0.0.0-dev/src/Bench.enso
+++ b/distribution/lib/Standard/Test/0.0.0-dev/src/Bench.enso
@@ -55,9 +55,9 @@ type Bench_Builder
     group : Text -> Bench_Options -> (Group_Builder -> Any) -> Any
     group self (name:Text) (configuration:Bench_Options) fn =
         validate_name name
-        b = Vector.new_builder
-        fn (Group_Builder.Impl b)
-        self.builder.append <| Bench.Group name configuration b.to_vector
+        group = Vector.build b->
+            fn (Group_Builder.Impl b)
+        self.builder.append <| Bench.Group name configuration group

 ## Builder to create a group of benchmarks.
 type Group_Builder
@@ -87,9 +87,8 @@ type Bench
     ## Construct a Bench object.
     build : (Bench_Builder -> Any) -> Bench
     build fn =
-        b = Vector.new_builder
-        fn (Bench_Builder.Impl b)
-        groups_vec = b.to_vector
+        groups_vec = Vector.build b->
+            fn (Bench_Builder.Impl b)
         Bench.All groups_vec . validate

[GregoryTravis]

I examined another, more traditional test (org_enso_benchmarks_generated_Column_Arithmetic_1000000_Multiply_Overflowing) and could not reproduce any difference between the same two revisions (for May 7 and 10).

At this point I also noticed that while many benchmarks show a regression around May 9, most of these regressions have reverted in the most recent sample, like this:

(But possilby this is an issue of incomplete data collection.)

[GregoryTravis]

It does seem to have returned to normal.