BEAM vs Erjang benchmarks

[Tony Arcieri]

Hello. I'm writing a blog post where I'm mentioning Erjang, specifically in regard to how it uses a shared heap (or rather, the JVM does). Are there any recent benchmarks of BEAM vs Erjang performance, specifically in regard to things which would benefit from a shared heap? (e.g. message passing performance)

--
Tony Arcieri

[Erik Søe Sørensen]

2011/6/7 Tony Arcieri <tony.a...@gmail.com>

Hello. I'm writing a blog post where I'm mentioning Erjang, specifically in regard to how it uses a shared heap (or rather, the JVM does). Are there any recent benchmarks of BEAM vs Erjang performance, specifically in regard to things which would benefit from a shared heap? (e.g. message passing performance)

Sorry about the late reply -
We have a CI server which also runs some performance tests - specifically, it runs the estone suite.

You asked specifically about message passing, and I'll address that, but first let me tell how the overall performance is...:

Overall, Erjang is performing slightly better than BEAM - as measured by the "EStones" number, which is a weighted average of the results of a number of microbenchmarks.

However, there are great variations in Erjang's performance among the different tests.

Where Erjang shines:
 * Message passing of huge messages
 * Floating-point arithmetics

Where Erjang has significant problems at the moment, performance-wise:
 * Scheduling / context switches (this includes timers)
 * Binary handling

For FP calculations, Erjang is >10x faster than Erlang. I have no idea why Erlang would be that sluggish; after all, BEAM bytecode has special instructions for FP operations.
I have plans for the scheduler, but hadn't had time to do something about them. (Recently I was going to look at it, but ended up fixing the >=R14B01 file subsystem problem instead...)
As for binaries - can't remember what the issue is there, if it is known, but it's twice as fast as it used to be - something was done about it shortly after the CI server started making performance graphs.

And, finally, message passing and inter-heap copying:
The EStone suite contains three message passing benchmarks: for "small", "medium" and "huge" messages.
The sizes of the messages (as measured by erts_debug:flat_size/1) are 3, 27 and 990 words, respectively.

Erjang's score on these benchmarks, relative to BEAM's, are:
 * Small: 40-45%
 * Medium: ~70%
 * Huge: ~800%

So - as expected - the larger the message, the larger the gain.
The horrible numbers for the "small" and "medium" benchmarks are (afaik) due to scheduler problems.

Regards,
Erik Søe Sørensen