Yonggang Che, Chuanfu Xu, Pingjing Lu, and Zhenghua Wang
Parallel simulation, trace overhead, localized trace storage, parallelI/O, performance
Trace-driven simulation is widely used in performance evaluation of computer architectures. Efficient trace accesses are crucial for trace-driven parallel simulators, as trace size grows rapidly with the number of target processors. This article presents a method named LPIO (Localized Parallel trace I/O) in the context of the BigSim parallel simulator. It utilizes the local disks of the host machines to store traces generated in the emulation phase. It proposes a task-local mapping algorithm that maps each simulation process to the node with the traces it needs. Hence trace I/O operations are performed in a scalable parallel manner with the support of all host nodes. Another benefit is that trace accesses are localized, avoiding data transfer through the interconnection network. Experiments with four parallel benchmarks on a SMP-cluster show that LPIO outperforms NFS (Network File System)-based method by 2.8% to 46.8% in the emulation phase. It also improves simulation performance up to 7.0% in the simulation phase. LPIO changes little of the simulator code or host-machine configuration and is easy to implement.
Important Links:
Go Back
