On Sun, Apr 8, 2012 at 00:38, Jann Horn <[email protected]> wrote:
> Hello,
> I've been using a little ad-filtering proxy I built using node for some
> time now, and today I had a look at its CPU usage while flash was
> prefetching a very large youtube video. Well, it was relatively high, I
> think, but the only thing it should be doing at that time is using
> the .pipe() method to pipe from one HTTP stream to another one.
>
> Well, I was surprised to see that node was wasting ~2.5% of its CPU
> usage on v8::String::New calls in node::MakeCallback, and that
> node::MakeCallback in total has a CPU overhead that amounts to ~7% of
> the total CPU time node uses in this case. Shouldn't it at least be
> possible to cache the symbol string? And couldn't node maybe also avoid
> calling v8::Object::Get each time it wants to call into userland (~3.6%
> )?
>
> Are these things right or did I make some huge mistake? If they are
> right, would it be possible to make this stuff a bit faster? Reducing
> the CPU usage by 5% seems like a relatively big speed improvement to me.
>
> Node version: 0.6.14
>
> Call graph centered on MakeCallback:
> http://thejh.github.com/callgraph.png
>
> Raw callgrind output:
> http://thejh.github.com/callgrind.out.19701

Yes, it's a known issue and something we'll address in v0.8, probably
by scrapping MakeCallback() and moving back to Persistent<T> handles.

In case you're interested, [1] is a branch where I did some
performance benchmarking. It turns out that MakeCallback() and V8's
thread-local storage functions (used to store the Isolate handle) are
the two most expensive functions.

The overhead is not terribly bad, just... mildly bad. In my tests
roughly 5% of total user CPU time is spent in those two functions.

[1] https://github.com/bnoordhuis/node/compare/benchmark

Reply via email to