[rvm-research] Replay compilation

[Dingwen Yuan]

Hi All,

I have a question to the evaluation with the help of "replay compilation". As written in the [1],
"During the replay phase, a benchmark is iterated twice within a single VM invocation. The first iteration includes the compilation overhead according to the compilation plan - this is referred to as mix run. The second iteration then is the timing run with adaptive (re)compilation turned off - this is called the stable run."
My question is whether we need to do something to "turn off the adaptive recompilation", for example by adding a line of code in the MMTkCallback if DaCapo is used, or we simply do nothing except setting -X:aos:enable_replay_compile=true and providing the advice files from the beginning.

Many thanks,

Dingwen

TU Darmstadt



[1] Andy Georges, Lieven Eeckhout, and Dries Buytaert. Java performance evaluation through rigorous replay compilation. In OOPSLA 2008 [OOPSLA08], 2008, pages 367-384.
------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now.  http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@...
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

[Andy Georges]

Hi Dingwen,

On Sep 9, 2009, at 17:28 PM, Dingwen Yuan wrote:

> Hi All,
>
> I have a question to the evaluation with the help of "replay  
> compilation". As written in the [1],
> "During the replay phase, a benchmark is iterated twice within a  
> single VM invocation. The first iteration includes the compilation  
> overhead according to the compilation plan - this is referred to as  
> mix run. The second iteration then is the timing run with adaptive  
> (re)compilation turned off - this is called the stable run."
> My question is whether we need to do something to "turn off the  
> adaptive recompilation", for example by adding a line of code in the  
> MMTkCallback if DaCapo is used, or we simply do nothing except  
> setting -X:aos:enable_replay_compile=true and providing the advice  
> files from the beginning.

The user guide points out the various options you need to provide (http://jikesrvm.org/Experimental+Guidelines 
):

During replay, you should use the following command line options

-X:aos:enable_replay_compile=true
-X:vm:edgeCounterFile=my_edge_counter_file
-X:aos:cafi=my_compiler_advice_file
-X:aos:dcfi=my_dynamic_call_graph_file

This of course assumes you recorded the edge counter data, the actual  
advice data and the dynamic call graph data. So your latter idea was  
the correct one. You need not hack the MMTKCallback class to tell the  
RVM something in between iterations. You should however make sure that  
the heap is fully collected prior to doing the measurement run.

-- Andy



------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now.  http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@...
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

[Dingwen Yuan]

Hi Andy,

Thanks, now I got it.

regards,

Dingwen

2009/9/10 Andy Georges <andy.georges@...>

Hi Dingwen,

On Sep 9, 2009, at 17:28 PM, Dingwen Yuan wrote:

> Hi All,
>
> I have a question to the evaluation with the help of "replay
> compilation". As written in the [1],
> "During the replay phase, a benchmark is iterated twice within a
> single VM invocation. The first iteration includes the compilation
> overhead according to the compilation plan - this is referred to as
> mix run. The second iteration then is the timing run with adaptive
> (re)compilation turned off - this is called the stable run."
> My question is whether we need to do something to "turn off the
> adaptive recompilation", for example by adding a line of code in the
> MMTkCallback if DaCapo is used, or we simply do nothing except
> setting -X:aos:enable_replay_compile=true and providing the advice
> files from the beginning.

The user guide points out the various options you need to provide (http://jikesrvm.org/Experimental+Guidelines
):

During replay, you should use the following command line options

-X:aos:enable_replay_compile=true

-X:vm:edgeCounterFile=my_edge_counter_file
-X:aos:cafi=my_compiler_advice_file
-X:aos:dcfi=my_dynamic_call_graph_file

This of course assumes you recorded the edge counter data, the actual
advice data and the dynamic call graph data. So your latter idea was
the correct one. You need not hack the MMTKCallback class to tell the
RVM something in between iterations. You should however make sure that
the heap is fully collected prior to doing the measurement run.

-- Andy



------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now.  http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@...
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now.  http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@...
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers