Final submission

[merykitty]

Some small improvements:

• It is more performant to use branches than to use min/max
• Do 2 independent iterations to improve IPC

Some that I have tried:

• Do 4 iterations and use vector instructions to parse all of them at once: LongVector multiplication codegen is horrible, assembling a vector from its elements is expensive, also the constants need to be materialized for each iteration unless the compiler can inline the whole thing.
• Spawning a separate process to hide the cost of unmapping: It is likely that the cost of initializing the VM is higher than the cost of unmap, I did not see improvements.
• Some trickeries to avoid Unsafe: Spawning a new segment each time introduces regression, most likely due to failures to inline even with that threshold of inlining, and using a Long.MAX_VALUE-size segment introduces additional bound checks.

Check List:

• You have run ./mvnw verify and the project builds successfully
• Tests pass (./test.sh <username> shows no differences between expected and actual outputs)
• All formatting changes by the build are committed
• Your launch script is named calculate_average_<username>.sh (make sure to match casing of your GH user name) and is executable
• Output matches that of calculate_average_baseline.sh
• For new entries, or after substantial changes: When implementing custom hash structures, please point to where you deal with hash collisions (line number)

• Execution time: Cannot tell, my laptop is not really reliable for benchmarking, but the elapsed time is around 20s.
• Execution time of the previous implementation: Around 26s.

[tarsa]

• Spawning a separate process to hide the cost of unmapping: It is likely that the cost of initializing the VM is higher than the cost of unmap, I did not see improvements.

it seems there's another trick to try:

1brc/src/main/java/dev/morling/onebrc/CalculateAverage_serkan_ozal.java

Lines 171 to 176 in 7f0e517

	if (CLOSE_STDOUT_ON_RESULT) {
	// After printing result, close stdout.
	// So parent process can complete without waiting this process completed.
	// Saves a few hundred milliseconds caused by unmap.
	System.out.close();
	}

along with

1brc/calculate_average_serkan-ozal.sh

Lines 31 to 35 in 7f0e517

	#echo "Process started at $(date +%s%N | cut -b1-13)"
	eval "exec 3< <({ $CONFIGS java $JAVA_OPTS --class-path target/average-1.0.0-SNAPSHOT.jar dev.morling.onebrc.CalculateAverage_serkan_ozal; })"
	read <&3 result
	echo -e "$result"
	#echo "Process finished at $(date +%s%N | cut -b1-13)"

[gunnarmorling]

Nice!

#	Result (m:s.ms)	Implementation	JDK	Submitter	Notes
	00:02.367	link	21.0.1-open	Quan Anh Mai	uses Unsafe
	00:03.210	link	21.0.1-open	Quan Anh Mai

[gunnarmorling]

Hey @merykitty, couldn't find another way to reach out to you, so commenting here: would you perhaps be interested in teaming up with Thomas, Roy, Elliot, and me for doing a live stream on #1BRC tomorrow with Nicolai Parlog: https://twitter.com/gunnarmorling/status/1752288991909478611. As one of the fastest JVM-based implementations, it would be great to have your perspective on the challenge. Let me know :)

[merykitty]

@gunnarmorling I would be more than happy to join. Can you give me more information regarding the subjects or what can I contribute to the conversation? My email is [email protected].

[gunnarmorling]

Perfect, I've added you to the DM chat with the other participants. Thanks a lot, looking forward to it!

…

Message ID: ***@***.***>