Recoding past applications to prepare for the future of supercomputing
In early 2013 ORNL will once again be in the race to have the fastest supercomputer in the world. The Cray XT5 Jaguar HPC system is currently receiving a transformational upgrade to add graphics processing units (GPUs) to augment its central processing units, leading up to its rebirth as a speedy Cray XK6 named Titan. Of course just as in a race, the team at ORNL has several hurdles it must be clear before this hybrid powerhouse can cross the finish line.
One of the lead scientific athletes jumping the hurdles is Bronson Messer, senior research and development staff member and acting leader of the Scientific Computing group at the National Center for Computational Sciences at ORNL. On August 14 in Washington, DC, he addressed the Advanced Scientific Computing Advisory Committee—the external board charged with advising the DOE Office of Science’s Center for Advanced Scientific Computing Research on scientific and technical issues—on how to be ready for “first light” on Titan.
“The objective was to let them know what we plan to do sciencewise as soon as Titan is ready for business,” Messer says.
Because Titan will be incorporating accelerated GPUs—“kind of like the ones in your PS3,” Messer says—the math or language in which the scientific application codes talk to the processors has changed and needs to be restructured. He and his team have been working to rewrite the codes so that they can properly allocate specific jobs within the application to their corresponding destinations within the newly added 16-core AMD processors.
Messer and his colleagues at the Center for Accelerated Application Readiness, as well as collaborators from NVIDIA, the Swiss National Supercomputing Centre, and the National Center for Supercomputing Applications have already come up with about a dozen GPU-capable applications that will be ready to run day-one simulations on Titan. Once this computational giant is complete, the team will become the pioneers in the next generation of accelerated GPU simulation users.
“We are going to run simulations that take a major fraction if not all of the machine, all at once, to perform simulations that have never been done before,” Messer says. —by Jeremy Rumsey