With the advent of petascale computing era, a typical petascale machine is usually installed with hundreds of thousands of processors. To predict the performance of existing applications running on such a machine, a simulation-based approach faces various unprecedented challenges. One of them is the memory constraint posed by emulating the petascale application that uses petabytes of memory on an existing parallel machine that typically only has hundreds of terabytes of memory or less. With an attempt to address this challenge, we present a preliminary investigation of utilizing the out-of-core concept in the context of an existing performance prediction framework BigSim. We have designed three basic out-of-core schemes and explored the optimizations that could be applied to the basic schemes. The additional theoretical analysis models both the performance of the emulation with and without out-of-core support, points out the very importance of the optimizations for the basic out-of-core schemes and offers suggestion to its implementation. By resolving multiple issues across the software stack, we implemented a prototype of the basic out-of-core support and integrated it into the emulation component of BigSim. The experimental evaluation shows the prototype works correctly for both Charm++ and MPI applications, and a reasonable emulation performance regarding a MPI kernel benchmark.
