benchmark improvements

[mranney]

There is a lot of funny business with benchmarking node programs due to V8's very clever GC and recompilation behavior. To work around this, I suggest running each test several times so that the total runtime ends up being around a minute per type of test. Varying the order might also yield different results.

An added benefit of smoothing out the variance by running longer is that you can capture some stacks with DTrace and see where the time is actually going.

[anandsuresh]

Sounds good to me. Hopefully I should be able to get that done over the upcoming weekend.

On Dec 2, 2013, at 6:20 PM, Matt Ranney [email protected] wrote:

benefit of smoothing out the variance by running longer is that you can capture some stacks with DTrace and see where the time is actually going

[anandsuresh]

For a start, I’ve modified the code to use buffers instead of strings. That little change has made quite a bit of difference in the GC interference. Running the benchmarks with —trace_gc reveals that it barely runs the GC during the benchmarks.

I do like the idea of running vFlame and dTrace to get more insights into where its spending the most time.

On Dec 2, 2013, at 6:21 PM, Anand Suresh [email protected] wrote:

Sounds good to me. Hopefully I should be able to get that done over the upcoming weekend.

On Dec 2, 2013, at 6:20 PM, Matt Ranney [email protected] wrote:

benefit of smoothing out the variance by running longer is that you can capture some stacks with DTrace and see where the time is actually going