headius

This is an experiment to pre-train and AOT enough of JRuby to get startup time for "gem list" down to a reasonable speed. I've tried various combintions and so far this seems to be the best, but the timings are highly variable.

Step 1: Perform a gem list while dumping IR. This saves the compiled IR for all loaded files to disk, to skip parsing and compiling in future runs.

jruby --dev -Xir.writing -S gem list > /dev/null

Step 2: Train AOT by running the following script repeatedly with the given JRuby arguments. Requires a small JRuby patch because IR reading breaks for any script has not been written (like the training script).

headius

Native Memory Tracking:

(Omitting categories weighting less than 1KB)

Total: reserved=1341322KB, committed=143578KB
       malloc: 30682KB #118112, peak=33492KB #117863
       mmap:   reserved=1310640KB, committed=112896KB

-                 Java Heap (reserved=57344KB, committed=32768KB)
                            (mmap: reserved=57344KB, committed=32768KB, at peak)

headius

#
# A fatal error has been detected by the Java Runtime Environment:
#
#  SIGBUS (0xa) at pc=0x0000000103e8e604, pid=4379, tid=35331
#
# JRE version: OpenJDK Runtime Environment (23.0+1) (build 23-valhalla+1-90)
# Java VM: OpenJDK 64-Bit Server VM (23-valhalla+1-90, mixed mode, sharing, tiered, compressed oops, compressed class ptrs, g1 gc, bsd-aarch64)
# Problematic frame:
# V  [libjvm.dylib+0x9ce604]  SystemDictionaryShared::ArchiveInfo::lookup_lambda_proxy_class(LambdaProxyClassKey*)+0x5c
#

headius

diff --git a/core/src/main/java/org/jruby/RubyBasicObject.java b/core/src/main/java/org/jruby/RubyBasicObject.java
index 4edab9b18a..985f2d1911 100644
--- a/core/src/main/java/org/jruby/RubyBasicObject.java
+++ b/core/src/main/java/org/jruby/RubyBasicObject.java
@@ -1791,8 +1791,8 @@ public class RubyBasicObject implements Cloneable, IRubyObject, Serializable, Co
      * it possible to emulate both instance_eval and instance_exec
      * with this implementation.
      */
-    protected IRubyObject yieldUnder(final ThreadContext context, RubyModule under, IRubyObject[] args, Block block, EvalType evalType) {
-        context.preExecuteUnder(this, under, block);

headius

Comparing the run times for the most intensive job in our CI suite. JRuby 9.4 and 10 are both on Java 21 running equivalent settings. JRuby 10 can automatically utilize OpenJDK's AppCDS to improve startup time, and has many other optimizations for straight-line performance.

"Force JIT" here means JIT is enabled with a threshold of zero, causing nearly all methods and blocks to fully optimize and compile to JVM bytecode before running. JRuby 10 includes reductions in the overhead of the JIT and more are coming.

InvokeDynamic also can slow startup even as it improves execution performance. We have worked to reduce the overhead of InvokeDynamic use in JRuby while also utilizing it more effectively.

JRuby 9.4, force JIT, with full invokedynamic optimization.

Finished tests in 1053.596852s, 4.4125 tests/s, 1756.1271 assertions/s.

headius

Baseline startup

Base startup numbers with and without CDS (--nocache disables CDS), with and without --dev (simpler JVM JIT and no JRuby JIT)

[] jruby $ time jruby --nocache -e 1      
jruby --nocache -e 1  2.48s user 0.17s system 148% cpu 1.777 total
[] jruby $ time jruby --nocache -e 1
jruby --nocache -e 1  2.37s user 0.13s system 185% cpu 1.340 total
[] jruby $ time jruby --nocache -e 1

headius

JRuby without any flags

JRuby's startup without any flags is not unreasonable, but we're starting from a bad place.

jruby -e 1  2.57s user 0.17s system 166% cpu 1.648 total
[] jruby $ time jruby -e 1
jruby -e 1  2.46s user 0.13s system 183% cpu 1.407 total
[] jruby $ time jruby -e 1
jruby -e 1  2.48s user 0.13s system 179% cpu 1.455 total

headius

These are resources suggested by my Java friends online.

Free online resources

• W3Schools Java tutorial A nice high-level tutorial of basic Java features.
• SAMS Teach Yourself Java There's multiple editions for different Java versions, so later editions may use features the robobrain doesn't support.
• Modern Java Recommended by someone who used this to teach his kids.

headius

private IRubyObject complexPowForSpecialAngle(ThreadContext context, IRubyObject other) {
    if (!(other instanceof RubyInteger integer)) {
        return UNDEF;
    }

    IRubyObject x = UNDEF;
    int dir;
    if (f_zero_p(context, image)) {
        x = real;
        dir = 0;

headius

~/work/jruby $ ~/amazon-corretto-25.0.0.1.1-linux-x64/bin/java -version
openjdk version "25" 2024-12-10
OpenJDK Runtime Environment Corretto-25.0.0.1.1 (build 25+1-Nightly)
OpenJDK 64-Bit Server VM Corretto-25.0.0.1.1 (build 25+1-Nightly, mixed mode, sharing)

~/work/jruby $ JAVA_OPTS="-XX:+UseParallelGC" jruby bench/bench_csv.rb 
WARNING: A terminally deprecated method in sun.misc.Unsafe has been called
WARNING: sun.misc.Unsafe::arrayBaseOffset has been called by org.jruby.util.StringSupport (file:/home/headius/work/jruby/lib/jruby.jar)
WARNING: Please consider reporting this to the maintainers of class org.jruby.util.StringSupport
WARNING: sun.misc.Unsafe::arrayBaseOffset will be removed in a future release