[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Thong (Tum) Nguyen
Tue, 18 May 2004 14:39:28 +1200
I just wanted to give everyone an update on the things I've been working on
in the past couple of weeks. Hopefully this message will help make the
developers list look busy and attract more developers :D!
Monitors are now much faster than they were in 0.1. Thanks to ideas and
encouragement from Russell Stuart and a few optimisations and
simplifications of the algorithm, the code runs several orders of magnitude
faster than the monitors in 0.1 (by my tests, over 200 times faster). This
should speed up the class library in general.
Here's some brief very unscientific benchmarks for the new monitors against
a commercial .NET VM.
Test PNET Commercial VM
Uncontested no lock 33,695,060 313,115,470
Uncontested 6,612,628 30,750,238
Contested recursive 6,323,352 29,521,962
Contested-2-threads 6,669,712 25,270,728
The test is a loop that increments a counter. The no lock test is for
PNET is about 1/5th the speed when counting with locks compared to when
The commercial VM is about 1/10th the speed when counting with locks
compared to when counting without.
When not using locks, PNET is 1/10th the speed of the commercial VM.
When using locks, PNET is 1/5th the speed of the commercial VM.
Probably doesn't mean much since we're comparing a JIT to an unroller but I
think it does indicate that PNET monitors show promise. This is especially
true when you consider that MS.NET monitors are broken. For example, MS.NET
won't throw a ThreadStateException when you exit a monitor without entering
PNET supports two methods of associating monitors with objects. The
standard method stores a pointer to the monitor in the object header. The
"thin-lock" method uses a hashtable to associate the monitor with the
The standard method is faster but requires an extra word for every object
even if the object is never used for synchronization. The thin-lock method
is slower but doesn't require an extra word for every object so is better
on platforms with limited memory or with programs that create thousands
(or millions) of tiny objects or with programs that don't use make much
use of synchronization (remember: pnetlib uses synchronization).
The monitor enter/exit algorithms are designed to eliminate context switches
into kernel mode by avoiding "fat" OS-locks (therefore avoiding system
calls) and as I mentioned above, are now much faster than they were before.
You can enable thin-locks by configuring pnet with the full-tl profile
I think Portable.NET is currently the only .NET VM to offer thin-locks.
Thin-lock support is inline with one of Portable.NET's biggest advantages:
its ability to run on small devices with limited memory.
Lots of improvements here. The threading stuff in pnet/support is
stabilising nicely -- lots of sneaky race conditions have been eliminated
thanks to tests from Russell. Interrupt and Abort support has been touched
up and they are now in a more usable state ;-). Interlocked functions have
been movied to the libsupport and the x86 versions of these functions have
been optimised using assembly. It would help if some volunteers could come
along and write PPC/MIPs/SPARC/etc versions of these functions.
You can now invoke methods with out/ref parameters and get those parameters
back :-). The out/ref parameters will be passed back through the parameter
array. One interesting thing I learnt is that .NET allows you to pass nulls
as arguments to functions that require valuetypes (or valuetypes byref).
The nulls are converted by the runtime into empty value-types of the
corresponding type. AFAIK, this is undocumented and mono throws an
NullReferenceException instead. I've made PNET's behaviour mirror MS.NET.
Asynchronous method calls (via delegates) now work -- including the ability
to retrieve out/ref parameters in the call to Delegate.EndInvoke. Out/ref
parameters will be passed back into the stack/fields just like with any
other method call.
Foundations for typed allocation have been laid (see include/il_gc.h).
Typed-allocation will allow us to decrease the chances that objects won't be
collected or finalized by letting us tell the GC which words within an
object contain pointers and which don't.
All the best,
- [DotGNU]Engine updates,
Thong (Tum) Nguyen <=