Vorticity structures in a rotating model star with convection in equilibrium in its unstable outer half, PPM simulation by Porter, Anderson, Habermann, and Woodward on two 64-processor SGI Origins at NCSA, 1997.
The simulation used two 64-processor systems interconnected by 100 megabyte per second HiPPI channels. NCSA plans to upgrade this system within the year and to upgrade the interconnection network to the 800 megabyte per second CrayLink.
For this simulation Porter, Anderson, and Habermann developed a new version of the LCSE's PPM gas dynamics code that employs a novel strategy for overcoming latency in parallel applications. Latency refers to the time it takes to retrieve data from memory. In large multiprocessor computing systems, such as the Origin2000, each group of processors that shares a common memory is simultaneously advancing a different portion of the problem. This means of attacking the problem in parallel is how the computer attains speed and efficiency. Latency becomes an issue when the processor groups (there were two groups of 64 processors used at NCSA) must retrieve information from other processor groups in order for the calculations to continue. This communication between groups of processors introduces latency.
The LCSE team's new PPM code prevents latency from slowing a computation because it overlaps the computation with the exchange of information among the other processor groups. Rather than eliminating latency, it hides it. Hiding latency becomes easier as the size of the problem increases because the amount of time spent computing increases faster than does the time spent exchanging data between processor groups. In large problems, such as solar convection, the time spent computing overwhelmingly exceeds the time spent exchanging data.
Velocity structures in a rotating model star with convection in equilibrium in its unstable outer half, PPM simulation by Porter, Anderson, Habermann, and Woodward on two 64-processor SGI Origins at NCSA, 1997.
