Vector_Operations.Partition benchmark results in an infinite deopt loop
#11437
Assignees
Labels
Comments
|
Another example of infinite deopt loop is in https://github.com/enso-org/enso/actions/runs/11564721429/job/32190462547#step:7:11096 in benchmark
|
|
Can be reproduced simply by running the benchmark directly in pure Enso via: |
|
Running: reveal a lot of |
|
Apply this patch: diff --git a/engine/runtime/src/main/java/org/enso/interpreter/runtime/data/vector/ArrayBuilder.java b/engine/runtime/src/main/java/org/enso/interpreter/runtime/data/vector/ArrayBuilder.java
index 5da2ece85b..b8e2820345 100644
--- a/engine/runtime/src/main/java/org/enso/interpreter/runtime/data/vector/ArrayBuilder.java
+++ b/engine/runtime/src/main/java/org/enso/interpreter/runtime/data/vector/ArrayBuilder.java
@@ -13,9 +13,12 @@ import com.oracle.truffle.api.nodes.Node;
import java.util.Arrays;
import org.enso.interpreter.runtime.data.EnsoObject;
import org.enso.interpreter.runtime.warning.WarningsLibrary;
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
@ExportLibrary(InteropLibrary.class)
final class ArrayBuilder extends EnsoObject {
+ private static final Logger LOGGER = LoggerFactory.getLogger(ArrayBuilder.class);
private static final String[] MEMBERS =
new String[] {"isEmpty", "add", "appendTo", "get", "getSize", "toArray"};
private final int initialCapacity;
@@ -79,6 +82,8 @@ final class ArrayBuilder extends EnsoObject {
if (e instanceof Double d) {
if (size == doubleArray.length) {
CompilerDirectives.transferToInterpreter();
+ LOGGER.debug("[ArrayBuilder@{}] double array capacity reached: {}, resizing to: {}",
+ Integer.toHexString(System.identityHashCode(this)), size, size * 2);
primitiveArray = doubleArray = Arrays.copyOf(doubleArray, size * 2);
}
doubleArray[size++] = d;
diff --git a/test/Benchmarks/src/Vector/Operations.enso b/test/Benchmarks/src/Vector/Operations.enso
index 58f9404add..7f1946209f 100644
--- a/test/Benchmarks/src/Vector/Operations.enso
+++ b/test/Benchmarks/src/Vector/Operations.enso
@@ -9,7 +9,7 @@ import project.Vector.Utils
polyglot java import java.util.Random as Java_Random
-options = Bench.options . set_warmup (Bench.phase_conf 5 5) . set_measure (Bench.phase_conf 1 5)
+options = Bench.options . set_warmup (Bench.phase_conf 3 3) . set_measure (Bench.phase_conf 1 1)
collect_benches = Bench.build builder->And run: Gives the output: TL;DR; There are a lot of ArrayBuilders with a primitive double array that are resized up to the capacity of 10^6. When an ArrayBuilder is resized, there is |
|
Pavel Marek reports a new STANDUP for today (2025-02-25): Progress: - Revealed most probable cause of the indefinite deoptimizations in Vector.Partition benchmark:
|
|
Pavel Marek reports a new STANDUP for today (2025-02-26): Progress: - Discussion about engine/stdlib telemetry and logging.
|
This was referenced Feb 27, 2025
|
Pavel Marek reports a new STANDUP for today (2025-02-27): Progress: - Trying the approach with methods behind
|
github-project-automation
bot
moved this from 👁️ Code review
to 🟢 Accepted
in Issues Board
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
On 4898445, reproduce with:
The output is:
Note that because of the infinite deopt loop, the stddev is even bigger than score, so the result from the benchmark is not reliable.
Output on the CI: https://github.com/enso-org/enso/actions/runs/11564721429/job/32190462547#step:7:11505
The text was updated successfully, but these errors were encountered: