Re: [rvm-research] Using hardware performance monitors to understand the behaviour of java applications

[Steve B. <Ste...@an...>]

Ian,

It's nice to hear you have oprofile working.  I was responding to  
Andy's comments.   Laurence's libpfm work is a modern replacement for  
the code Andy refers to.

--Steve

PS I understand what oprofile is, what its virtues are (but that is  
not what Andy was using).  I also understand the status of the kernel  
support for perf_events (as it happens Paul MacKerras lives a few  
houses down the road and I'm well aware of what they're doing.)

On 31/05/2010, at 7:38 AM, Ian Rogers wrote:

> On 28 May 2010 17:21, Steve Blackburn <Ste...@an...>  
> wrote:
>> Two comments:
>>
>> 1.  libpfm4, which leverages the new linux kernel support for perf
>> events is the way to go, and it makes most of our previous work on
>> perf counters redundant.   We've just done a lot of fairly intense
>> work with it here (not Jikes RVM related), and it is indeed good.
>>
>> 2.  Laurence Hellyer has gone and implemented pretty much exactly
>> this.   I hope his work is folded into the head soon (I guess Daniel
>> and I particularly need to move on that front).   Anyway, the work is
>> available to you (Andy, Ankush, anyone else) to use now (http://jira.codehaus.org/browse/RVM-885
>> ).
>>
>> --Steve
>
> Currently the Linux kernel is trying to clean up and standardize
> performance counter implementations, this is a somewhat orthogonal
> issue as to how to interface with the counters. The libpfm4 approach
> is for someone to manually insert counters, for example around the
> complete execution of a thread. You can change your code and see what
> effect this has on overall execution time. This is a good way of
> getting performance numbers once you have a performant system. The
> OProfile approach uses sampling of counters and gives tools, which are
> built on by tools like AMD's code analyst, to drill down and find
> where execution time is being spent. You can identify hot spots in
> your code that are causing cache misses or executing many
> instructions. At the granularity of a thread these performance issues
> wouldn't be apparent. There is an OProfile to JVMTI bridge for
> drilling down into performance of off the shelf JVMs. MRP [1] provides
> this interface for Jikes RVM which doesn't support JVMTI.
>
> Regards,
> Ian
>
> [1] http://mrp.codehaus.org/
>
>> On 28/05/2010, at 11:19 PM, Andy Georges wrote:
>>
>>> Hi Ankush,
>>>
>>>
>>> On May 26, 2010, at 22:31 PM, Ankush Gupta wrote:
>>>
>>>> The paper "Using hardware performance monitors to understand the
>>>> behaviour of java applications" (http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.89.4848&rep=rep1&type=pdf
>>>> ) refers to monitoring the Hardware performance counters inside the
>>>> Jikes JVM. Also the paper "Bottleneck Analysis in Java Applications
>>>> using Hardware Performance Monitors" refers to monitoring the Java
>>>> application performance using the Hardware performance counters. Is
>>>> the work added to the existing Jikes RVM as patch or release and is
>>>> possible to repeat the experiments ?
>>>
>>> None of this code is readily available in the latest Jikes RVM or
>>> the svn head, but it should certainly be possible to repeat at least
>>> the experiments from our 2004 paper, as you need not much besides
>>> some instrumentation at the prologue and epilogue of a method and a
>>> way to read counter data (RDPMC) at these points. I do have our
>>> analysis scripts lying somewhere, but trust me, you do not want to
>>> go about figure them out, as they tend to be hacked together with
>>> the purpose of obtaining correct results quickly, rather than
>>> elegantly.
>>>
>>> -- Andy
>>> ------------------------------------------------------------------------------
>>>
>>> _______________________________________________
>>> Jikesrvm-researchers mailing list
>>> Jik...@li...
>>> https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers
>>
>>
>> ------------------------------------------------------------------------------
>>
>> _______________________________________________
>> Jikesrvm-researchers mailing list
>> Jik...@li...
>> https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers
>>
>
> ------------------------------------------------------------------------------
>
> _______________________________________________
> Jikesrvm-researchers mailing list
> Jik...@li...
> https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers