[erlang-questions] HiPE performance gain (Was: Re: [erlang-questions] Erlang static linked interpreter)

[Jesper Louis Andersen]

On Fri, Jan 21, 2011 at 20:50, Matthias Lang <matthias@REDACTED> wrote:

>
> I've assumed that what happens is that each new Erlang release is
> tuned for increasingly faster CPUs, while our hardware doesn't change
> as often. But I'm just guessing.
>

In my experience, the only way to combat performance regressions like
these is to plot the development branch against the time to complete
the benchmark suite. That way, the relative price of any change can be
quickly valuated. A page with sparklines usually suffices. I did one
for my Haskell/BitTorrent experiment:

http://jlouis.github.com/combinatorrent/

I've been thinking of using the performance counters in modern CPUs to
measure the amount of cache misses and plot these as well, but I never
came around to do it. I imagined using the linux 'perf' tool which
lives in the linux kernel tree under /tools/perf.


-- 
J.