Parallel Programming Talk #77 - Charles Leiserson

Welcome to Show 77 of Parallel Programming Talk originally broadcast on May 18th, 2010. On the show Clay talked with Prof. Charles Leiserson (Massachusetts Institute of Technology, Cilk Arts) about why programmers don't get good speed up in their parallel programs.  (Aaron was on vacation this week, so there's not much motion on the video version of the show.)



Download Link - MP4 Video File (Large)

Download Link - MP3 Audio File (Small)

First The News:

Three Webinars About Multicore Parallelism- May 12, 19 & 26th.
James Reinders will be talking with fellow industry leaders and innovators:

    • 12 MAY - Herb Sutter: "Five Years Since Free Lunches: Making Use of Multicore Parallelism"

    • 19 MAY - Steve Teixera: "Intel Parallel Studio and Microsoft Visual Studio: Using Abstractions to Prepare for Multicore"

    • 26 MAY (8am PDT) - David Mackay: "Intel Parallel Studio: Creating Fast, Reliable Parallel Programs"



If you missed any of these during their live broadcast, you can still register and watch them on demand.  However, don't wait too long.  The webinar presentations won't be available forever.

UPCRC Illinois Summer School is still accepting applications for participation, but there are a limited number of spots available(as I write these words).  The course will be held July 19-23, 2010 at the University of Illinois at Urbana-Champaign. Prerequisites for the summer school include solid programming experience (C, C++, C# or Java languages) and a demonstrated interest in applying multicore programming to academic or professional pursuits. If you are unable to attend in person at the Urbana campus, you can attend online.

Clay, without a "second banana" to banter with during the show, forgot to give the email address of the show. If you have questions you'd like to see up discuss, ideas for show topics or just want to send fan mail....
Send Email to parallelprogrammingtalk@intel.com

On the show today:

Dr. Charles LeisersonDr. Charles Leiserson, co-author of Introduction to Algorithms (MIT Press), discussed the four pitfalls programmers can run up against that will result in less than optimal speedup from parallel programs.  These problems are:

    • Insufficient parallelism in the application

    • Task management overheads

    • Memory bandwidth

    • Contention and true/false sharing



Charles gave examples of these problems along with tools and methods to detect or avoid them.  To find out how to side-step these problems in your won parallel coding efforts, listen to the advice of this eminent authority on algorithms with over 35 years of experience in parallel computation.

Some Useful Links:

Cilk SDK in the whatif.intel.com site 

(I can't find any Intel Cilk beta program site.  If you know where this is, please add a comment with a URL.)

Supercomputing Technologies Group (SuperTech)




Watch Parallel Programming Talk every Tuesday at 8:00AM PT.

有关编译器优化的更完整信息,请参阅优化通知