Abstract
Memory hierarchy has played a decisive role in the System-on-Chip (SoC) for providing necessary data bandwidth to maximize the system performance. In addition, memory subsystems with growing capacity and operating speed also account for a significant fraction of the energy consumption. However, our evaluation also shows that the simulation of memory sub-system can take up to 37% of the full-system simulation time. Therefore, an accurate and fast electronic system-level (ESL) model of DRAM controllers, has become vital to the design evaluation of the SoC. Among many high-level DRAM models nowadays, DRAMSim2 is a well-known cycle-based model which has been validated against the RTL design. In this thesis, we present an improved simulation model of DRAM controller based on DRAMSim2. First of all, an event-based extension of DRAMSim2 is addressed with the discussion and comparison with other existing approaches. To boost the simulation speed, the transaction speculation technique is proposed to minimizes cumbersome synchronization that dominates simulated time. Two variations has been implemented, i.e., the command-based and transaction-based approach. Comparing our timing model of two variations, while the command-based approach performs efficiently in general, the transaction-based approach is even superior in realistic cases with burst memory transactions. In addition, we also present the technique of the bank-state update skipping to reduce computation. Several experiments have been performed. Random memory transactions with different injection intervals, memory traces of SPEC 2006 benchmarks, and realistic GPU, ENC, DEC applications have been evaluated, showing the speedups up to 368.6 × against DRAMSim2. Our simulation model has also been integrated into an industrial SoC virtual platform and validated against its RTL implementation with the extended interface protocol and scheduling policies for the cycle accuracy. Finally, the experiment on the full-system ESL platform also shows that the proposed simulation model can improve the entire simulation time up to 4.4×, enabling the efficient full-system evaluation and optimization in the early design phase.