[erlang-questions] Strange optimization result

Thomas Lindgren thomasl_erlang@REDACTED
Mon Oct 22 17:37:26 CEST 2007


--- Kostis Sagonas <kostis@REDACTED> wrote:

> Thomas Lindgren wrote:
> > 
> > That's a 35% speedup just from sizing the heap
> right.
> > So it seems the internal VM heuristics for this
> could
> > be tuned or improved.
> 
> Yes, of course. The question is exactly how to
> satisfy
> all needs, especially since they are often
> contradictory. /.../

I think an online approach could be interesting here.
For example, record memory usage and GC
characteristics of processes spawned at one site (or
otherwise categorized somehow) and use that, along
with current memory use, to make educated (but
non-binding) guesses.

> Having such a
> repository of real applications that can be used as
> benchmarks
> is something that is currently lacking in the Erlang
> community.
> Contributions welcome.

The drawback of 'real code' is that it is often
proprietary and relies on environments that are
tedious to set up. It can be difficult to analyze and
optimize too, as readers looking in the Hipe archives
can see for themselves :-)

A few years ago, I collected and used a benchmark
suite of medium-size programs, based on OTP. These
exercised the compiler, mnesia, ASN1 (encode/decode
LDAPv2) and gen_tcp.

The advantage was that the code was free and
optimizing these apps would help everyone. A potential
disadvantage was that new versions appear regularly,
so the benchmark suite would have to evolve
appropriately and comparison between versions would be
iffy.

Alas, it didn't catch on at the time. But maybe the
right time has come. Today, I'd suggest adding yaws
and maybe YXA, the wide finder (of course!), and
suchlike to the mix.

> (*) The HiPE group can help in this -- this is a
> good MS Thesis.

I think so too.

Best,
Thomas


__________________________________________________
Do You Yahoo!?
Tired of spam?  Yahoo! Mail has the best spam protection around 
http://mail.yahoo.com 



More information about the erlang-questions mailing list