Manifold is a parallel simulation framework for multi-core systems. For full-system simulation, Manifold adopts the timing-directed simulation paradigm that separates the simulation into a functional front-end and a timing back-end. Components in the front-end perform functional simulation of the cores and send streams of instructions to the backend to simulate the timing behavior. In its current design, Manifold uses the QSim multi-core emulator as the frontend, which communicates with the back-end through network sockets. Experiments have shown that the latency of the socket communications has a significant impact on the overall simulation performance. This paper presents a novel method that attempts to hide the TCP/IP latency for the back-end by creating proxy processes as an intermediary between the front-end and the back-end. The proxies serve as clients to the QSim server in the front-end, and as servers to the back-end. They interact with the QSim server through sockets, while working with the back-end in a producer-consumer manner using shared memory segments. Experiments show that this method can completely hide the TCP/IP latency for the back-end. The back-end can always get its instructions from the shared memory without waiting for the QSim server. The overhead of getting inputs for the back-end simulation is reduced to almost zero. As confirmed by our experiments, this improvement causes some side effects that together lead to significant improvements in overall simulation performance. In testing of system models with up to 64 cores, we have achieved from 29% to 51% improvement in simulation performance.

An Efficient Front-End for Timing-Directed Parallel Simulation of Multi-Core System [paper]