Re: [DotGNU]Reminder - Weekly DotGNU Meeting

[Rhys Weatherley]

Gopal V wrote:

>         Is ilrun faster than mono JIT ? ... I heard an
> unverified comment that ilrun ran a particular library
> faster than mono-jit .. Is it possible ? ... How can
> an interpreter outperform a JIT ?..

Lies, damned lies, and benchmarks ...

Depends on a lot of factors.  What was the person measuring?
Time to run or time to load?  Was the app exceedingly complex,
requiring lots of method conversion overhead?

Load times can be quite impressive in pnet: I've had people
comment at demos that my engine is faster than Microsoft's
when they see programs load.  But it's mostly just smoke and
mirrors due to good algorithm choice.  (That, and I cheat
by pre-loading the app into the kernel's buffer cache prior
to the demo :-) ).

Complex apps may run a bit faster in pnet, because the overhead
of method conversion is smaller for CVM than for a JIT.  If the
app spends most of its time converting methods, and little time
actually executing the methods, then pnet may have a slight edge.

Method conversion is a trade-off: sometimes it gives a performance
boost and sometimes it doesn't.  I'm sure you'd be able to create
a program that ran faster in Mono's interpreter than pnet's, even
though logically you would think that a polymorphic interpreter
should be slower than a non-polymorphic one.

Measuring performance is difficult and depends upon a lot of factors.
Without access to the particular test case in question, it will be
difficult to see what is going on in this instance.

> I can't wait until pnet gets a full JIT ...

Me too.

Cheers,

Rhys.