Three new papers about parallelism in Mercury have been added to the papers page. Associated with each paper is a presentation that was given at ICLP 2011. The presentations are also available on the papers page.
I am the first author for two of these papers. One of them has been published in Theory and Practice of Logic Programming.
Thursday, August 18, 2011
Friday, August 12, 2011
Intel blogs and whitepapers have good titles but basically no content.
For instance http://software.intel.com/en-us/articles/finding-non-trivial-opportunities-for-parallelism-in-existing-serial-code-using-openmp/
This is titled "Finding Non-trivial Opportunities for Parallelism in Existing Serial Code using OpenMP" Sounds promising right? This is exactly the kind of work that I think is important (if I didn't I wouldn't be working on my thesis). However the article doesn't actually use OpenMP to find parallelism, only to exploit it. They use conventional profiling techniques to find places where parallelism might be useful but all the analysis must manually be done by the programmer. This analysis is: 1) finding the hotspots within the call graph 2) deciding if the hot spots can be parallelised 3) actually parallelising them.
The final step is done using OpenMP, so it's a parallelism annotation rather than re-writing the code completely, but this isn't new, we've seen OpenMP before. The rest of the article (like many of Intel's articles) is an advertisement for their software, in this case the different profilers that can be used. They also slip in references to other tools of theirs.
I'm probably not their intended audience, other people may be able to get more useful information from these articles such as "Use a profiler to work out where parallelism is likely to be a benefit" but that doesn't stop the article from being a giant advertisement which bothers me.
This is titled "Finding Non-trivial Opportunities for Parallelism in Existing Serial Code using OpenMP" Sounds promising right? This is exactly the kind of work that I think is important (if I didn't I wouldn't be working on my thesis). However the article doesn't actually use OpenMP to find parallelism, only to exploit it. They use conventional profiling techniques to find places where parallelism might be useful but all the analysis must manually be done by the programmer. This analysis is: 1) finding the hotspots within the call graph 2) deciding if the hot spots can be parallelised 3) actually parallelising them.
The final step is done using OpenMP, so it's a parallelism annotation rather than re-writing the code completely, but this isn't new, we've seen OpenMP before. The rest of the article (like many of Intel's articles) is an advertisement for their software, in this case the different profilers that can be used. They also slip in references to other tools of theirs.
I'm probably not their intended audience, other people may be able to get more useful information from these articles such as "Use a profiler to work out where parallelism is likely to be a benefit" but that doesn't stop the article from being a giant advertisement which bothers me.
Subscribe to:
Posts (Atom)