Over 20 members of the Gelato Federation are currently meeting in Pittsburgh, from Nov. 8-11, to review and exchange research advances for Linux on Intel's Itanium2 platform. A similar, ealier meeting Oct. 11-13 in Beijing attracted more than 25 Gelato Federation members.

The Gelato Federation is a worldwide consortium formed in March of 2002 by Hewlett-Packard (HP) and seven leading research institutions to foster the Itanium Linux platform.

Beijing meeting highlights

At the Beijing meetings, Mark K. Smith, Gelato's managing director, announced a grant program, sponsored by HP and Intel, for Gelato members. The program will supply a number of Itanium-equipped HP systems to Gelato members selected through a competitive proposal process. “This is a tremendous opportunity to catalyze some exciting new work and encourage some productive new collaborations among members,” Smith said.

Other discussions at the Beijing meeting focused on critically needed modifications to the gcc compiler (compiler improvements are a major focus of Gelato research), the expected arrival in 2005 of multi-core technology, and the adoption of Itanium driven by 64-bit extensions to x86.

Other topics included:

• Shin Yee Chung of Singapore's Institute for High-Performance Computing noted that with the growing disparity between processor and memory speed, use of cache to hide memory latency is critical to HPC. While cache can improve general performance, fine tuning to a specific cache configuration can greatly improve performance, Chung said. Unfortunately, fine-tuning by hand is very expensive. Cache-oblivious algorithms can fine-tune performance at runtime and improve cache performance dynamically, but they can also adversely affect overall performance due to function call overhead. Chung described research to minimize this overhead through cache-oblivious adaptive algorithms.
• An ORC-based OpenMP compiler was presented by the research team of Prof. Weimin Zheng of Tsinghua University. The team is identifying optimization opportunities to explore thread-level parallelism using OpenMP. The OpenMP binary was tuned for the IA64 architecture by integrating the OpenMP processing module with other ORC optimizations using performance data provided by pfmon and perflib. Results are promising, with up to 25 percent speedup per overhead cost.
• Bob Kidd of the University of Illinois (UIUC) announced the September 2004 beta release of OpenIMPACT, a high-performance, open-source C compiler developed for Linux-Itanium at UIUC. OpenIMPACT is engineered to support a gcc-like use model. Kidd shared some recent C++ support work and provided tips for using OpenIMPACT for more advanced optimizations.
• From CERN's openlab, Andreas Hirstius updated the Federation on development of the high-speed wide-area networking that will be required when CERN's Large Hadron Collider goes online in 2007. Anticipated data rates per experiment range from 200-1200MB/s, but because of large volumes, the data must be distributed to locations worldwide for analysis. The challenge is to set up reliable transoceanic 10Gb connections with these global partners. Hirstius focused his talk on how CERN is measuring stability, responsiveness, and recovery of the 10Gb network, as well as hardware and storage systems. Their Itanium cluster has demonstrated excellent stability and I/O capabilities for these networking tests.

Additional technical highlights from the Gelato meetings can be found at the Gelato portal.

 
This article was originally published on LinuxDevices.com and has been donated to the open source community by QuinStreet Inc. Please visit LinuxToday.com for up-to-date news and articles about Linux and open source.