willsewell / gc-latency-experiment Goto Github PK

Host Windows 10, 1909
Docker 2.2.0.3 (42716)
Docker is logged in.

$ make ocaml/results.txt
docker build -t gc-ocaml ocaml
Sending build context to Docker daemon  4.608kB
Step 1/10 : FROM ocaml/opam:debian-stable_ocaml-4.04.2
pull access denied for ocaml/opam, repository does not exist or may require 'docker login': denied: requested access to the resource is denied
mingw32-make: *** [Makefile:27: ocaml/results.txt] Error 1

Iterations should be timed manually instead of relying on the runtime system statistics. This will make it consistent with the other benchmarks.

cc @gasche

Normalise on just counting time at each iteration instead of trying to use the language runtime's monitoring flags, which are proving too fragile besides making the system more complex/fragile and harder to port to other languages.

I'm looking to use the Common Lisp version of the benchmark in a paper, detailing a parallel non-moving GC for SBCL. There wouldn't appear to be any licenses in this repository; would you be willing to add one please?

(cc: @q3cpma who wrote the CL code.)

Instead of doing N iterations with the same "channel" structure, I would like to do K rounds of N iterations, with each round allocating a new channel structure and forgetting about the new one. My intuition is that this would reveal latencies associated with deterministic reference-counting that are currently hidden from the benchmark while they may manifest themselves in realistic program. But I haven't actually experimented with it to check it; if there is no counting-collection pause time, it may not be worth the extra complexity.

I assume it is the case for other VM based languages, but simply passing the following args

-XX:+UseG1GC -Xmx512m -Xms512m -XX:MaxGCPauseMillis=10

lowered the max time from 182 ms to 28 ms

You need to understand the environment in which things are being run a little better to do benchmarks.

Messages should be stored in an array instead of a hash table to make it consistent with the other benchmarks.

Racket goes over my head unfortunately.

@gasche, you might be interested or know someone who would be...

Should I hire Senior Java Developer full-time to read what is written in java/results.txt?
Because it's 80 KB long.
Or it's OK to ignore the whole file and just read "Worst push time"?