Jean-Pierre Panziera describes in his paper how the Linpack NxN project for NASA's Columbia system served as an occasion to illustrate how shared memory parallelism can be the key to highly scalable code. He showed that an efficient shared-memory implementation might require a complete restructuring of the algorithm. In the case of Linpack NxN, both the SMPL code and the message-passing HPL code distribute the matrix across the different working threads.

However, the shared-memory approach accelerates the data exchange between threads, in particular during panel factorization, according to the research of Mr. Panziera. Shared memory is also fundamental in optimizing the inter-node communications. Shared-memory parallelism is often confused with loop-level parallelism, which does not scale very well in general. But if one takes a more global approach, shared-memory applications can be more scalable than their message-passing counterparts.

Mr. Panziera presented his winning paper in the morning session on Thursday, June 23.

John Wu in his paper explains that the Grid Collector combines an indexing technology and a Grid file management technology to make the analysis of high-energy physics data considerably faster and easier than using the existing analysis frameworks. For common analysis jobs where the required files are on disk, the Grid Collector can speed up the executions because it avoids reading unwanted events, as stated by Mr. Wu and his colleague writers. For analyses that involve files not already on disk, the Grid Collector automatically transfers the necessary files and avoids the tedious manual file management tasks.

As such, with the same computer resources, more analysis jobs can be performed with the Grid Collector. The Grid Collector can also make more computer resources available for analysis by making it easier to use the scattered computer resources outside of the computer centers. The authors note that the Grid Collector is designed as a plug-in for an existing analysis framework. This allows an on-going HENP experiment to easily boost their analysis capacity with a minimal amount of investment.

Mr. Wu presented his research in the session on "Data Management on Distributed Systems and Grids" on Friday, June 24.

Kenin Coloma in his winning study shows that one of the largest challenges in client-side caching in extremely large-scale environments is consistency and coherency. By handling a user-space cache, Mr. Coloma can offer applications much closer control over the client-side cache and scale the cache with the size of the compute resources.

In this way, cache data is shared among each compute node in an analague way to a traditional shared memory

machine. Coloma's approach to maintaining the integrity of the distributed cache turns out to be quite scalable and offers potentially sizable performance gains.

Mr. Coloma presented his results in the morning session on Friday, June 24.