[Mono-dev] System.Reflection.Assembly::Load Performance?

[Babbage Linden]

Also, there seems to be some caching going on here: I only see the slow down
the first time I call System.Reflection.Assembly::Load with a given set of
bytes. Subsequent calls with the same bytes being loaded in to the same
domain are much faster.

Are there any tools I can use to continue profiling from managed code back
in to the mono runtime?

Cheers,

Jim

2009/12/1 Babbage Linden <babbage at lindenlab.com>

> Hi all,
>
> I'm currently investigating a performance problem which is causing complex
> objects in Second Life containing many scripts to take a long time to load.
>
> We embed Mono in the Second Life simulator, so I turned on the profiler
> before initialising the runtime:
>
> mono_profiler_load("default:stat");
> mono_jit_init("root domain");
> mono_internal_call_init();
>
> After shutting down the simulator, the profiler reported the slow method as
> being System.Reflection.Assembly::Load(byte[]), with each call taking
> 19.961ms:
>
> Time(ms) Count   P/call(ms) Method name
> ########################
>  4052.135     203   19.961   System.Reflection.Assembly::Load(byte[])
>   Callers (with count) that contribute at least for 1%:
>          102  50 % LindenLab.SecondLife.Script::.ctor(string,byte[],long&)
>          101  49 % LindenLab.SecondLife.Script::Deserialize(byte[],byte[])
>
> Does this seem right? We're just passing an assembly as bytes to Mono and
> asking it to turn it in to an assembly, 19ms seems like a long time to do
> that.
>
> Are there any faster ways we could do this? In the example above an object
> with 200 scripts in it is taking 4s to load, which is much faster than with
> an LSL script.
>
> I get similar results using mono 1.2.6 (which we have embedded in Second
> Life) or revision 146428, which is pretty close to the tip of svn.
>
> Also, the profiler seems to output it's results the first time we unload a
> domain, is there a way to stop that happening, so we get the full output on
> shut down?
>
> (How should we be creating and destroying appdomains from unmanaged code:
> we're currently using ves_icall_System_AppDomain_InternalUnload from
> unmanaged code to unload domains, which is less than ideal...)
>
> Cheers,
>
> Jim
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.ximian.com/pipermail/mono-devel-list/attachments/20091201/4722308c/attachment.html