FastFwd: An efficient hardware acceleration technique for trace-driven network-on-chip simulation

We present an efficient emulation-based technique to accelerate architecture exploration of networks-on-chip (NoCs). The large design space of NoC along with its growing complexity that results in low simulation speeds on host machines have motivated the need for hardware accelerators for speeding up the simulation. For example, simulation of applications with real life problem sizes could take weeks on a host machine. FPGA acceleration is a promising strategy for speeding up NoC simulations by several orders of magnitude. However, it is required to simulate a few billion network transactions of the application during NoC exploration, and this could still take tens of minutes even with an FPGA-based emulator. With the increasing complexity of architectures and applications, reducing emulation time is a key concern. We propose a technique, FastFwd, to minimize emulation time by efficiently identifying and eliminating redundant cycles during a trace-based NoC simulation. We have studied the implications of the additional FPGA hardware required for implementing our technique. A naïve implementation could lead to poor scalability and increase the required DRAM bandwidth, both of which ultimately impact the emulation speed negatively. We propose a hierarchical controller architecture to resolve the scalability issue, and a compressed representation of traces for mitigating the increased DRAM bandwidth requirement. Our experiments with several benchmarks have shown that the FPGA emulation with our technique reduces the average emulation time by a factor of 2 when compared to a conventional emulation.