[parsec-users] The PARSEC Benchmark Suite fundamental properties.

Christian Bienia cbienia at CS.Princeton.EDU
Wed Aug 19 01:26:53 EDT 2009


Hi Muhammad abid,

 

To answer your questions:

 

1.)    The observation here is that some programs like dedup have multiple threads that execute different code. This is in contrast to, say, an HPC workload where each thread typically executes the same code. The progress of these threads depends on the timing mode. One thread of type A might do a lot of FPU computations, another thread of type B performs a lot of array scans and thus depends on low memory latency. What we do in PARSEC in these cases is that we run each pipeline stage with enough threads such that theoretically the entire CMP can work within only one pipeline stage. What you will then observe is that the number of active threads for each pipeline stage will stabilize so that just the right number of threads are active to maximize throughput. Suppose we take the example above and run it on an 8-way CMP with a slow FPU, then on average maybe 5 threads of type A will run and only 3 threads of type B. The other threads will be blocked most of the type. Suppose you run the same program on a version of the CMP with a blazingly fast FPU, then you might observe that on average only 3.5 threads of type A and 4.5 threads of type B are running at any time. What has happened is that the program has adapted to the details of the timing model.

2.)    Yes, the timing model affects characteristics such as lock contention but lock contention does not (much) affect characteristics such as the ones reported. For example, the working set sizes depend on the data structures and those don’t change with the timing model. But you’re right that there is likely to be some impact. There’s not much anybody can do about it, it’s simply part of the measurement inaccuracy.

 

Best,

Chris

 

 

From: parsec-users-bounces at lists.cs.princeton.edu [mailto:parsec-users-bounces at lists.cs.princeton.edu] On Behalf Of Muhammad abid Mughal
Sent: Wednesday, August 19, 2009 12:37 AM
To: PARSEC Users
Subject: [parsec-users] The PARSEC Benchmark Suite fundamental properties.

 

Hi Chris:

             Hope you are doing good. I got couple of questions :

 

1.In your research paper "The PARSEC Benchmark Suite: Characterization and Architectural Implications" you mentioned that " Some of its workloads can adapt to different timing models and can use threads to hide latencies. It is important to analyze these programs in the context of the whole system." i dont catch that. Could you please provide more details? Is it same idea as Fine-grained multithreading (different threads interleave to hide latencies.)

2. Fundamental properties of PARSEC workloads that are not affected by Timing Model are :

   a,Spatial/Temporal locality  b, miss /hit rate.

 

  but timing model do affect how different threads contend for a lock. This is also known as timing behavior of a program. Is n't it ?

 

Waiting keenly for yout kind response.

 

Thanks

Best regards,

Muhammad abid

 

 

 

 

 

 

  _____  

Get <http://sg.rd.yahoo.com/aa/mail/domainchoice/mail/signature/*http:/mail.promotions.yahoo.com/newdomains/aa/>  your new Email address! 
Grab the Email name you've always wanted before someone else does!

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.cs.princeton.edu/pipermail/parsec-users/attachments/20090819/257cebf0/attachment.html>


More information about the parsec-users mailing list