This will develop and validate SimFlex, a framework that applies simulation sampling and checkpointing techniques to enable fast, accurate and flexible simulation of large-scale systems. Based on preliminary experience, it is expected that simulation turnaround will be reduced by several orders of magnitude, while achieving high accuracy and confidence in the simulation results.To estimate performance, SimFlex systematically selects a large number of short sampling units from a program execution. SimFlex performs detailed cycle-accurate measurement only within each sampling unit. SimFlex provides a collection of complementary techniques, including checkpointing, to rapidly generate the architectural and microarchitectural state necessary to initialize the cycle-accurate models before each sampled measurement. SimFlex uses statistical sampling theory to determine the minimally sized sample necessary to achieve simulation results at a desired degree of accuracy. Overall, only an exceedingly small fraction of the complete execution is simulated in detail. Preliminary investigation of the SimFlex framework in the context of an out-of-order superscalar uniprocessor (e.g., Intel Pentium 4) has resulted in a simulator that can estimate performance for SPEC CPU2000 benchmarks with a worst-case error of 2.4% and an average error of 0.6% relative to full-benchmark simulation by measuring a sample of only 50 million instructions per benchmark. In contrast, full-benchmark simulations require simulating up to trillions of instructions. The project,will explore and evaluate the effectiveness of sampling simulation of large-scale multiprocessor systems. A successful outcome of this research will permit, for the first time, simulating the execution of realistic workloads on large-scale (e.g., 64-1024 nodes) system sizes. The SimFlex framework will be evaluated for large-scale system simulation in the context of a cycle-accurate distributed shared-memory multiprocessor simulator. The proposed simulation framework will be based on Virtutech's Simics, a substrate that enables full-system emulation, boots an entire off-the-shelf OS (e.g., RedHat Linux or Sun Solaris), and can execute/emulate unmodified commercial applications (e.g., database, web, and file servers). The porject wu`ill enable integrate simulation sampling and checkpointing techniques to accelerate the simulation performance of this large-scale full-system simulator. The resulting detailed system simulator will enable performance study using realistic workloads and accurately capture not only application behavior but also OS interactions.