LACross: Learning-Based Analytical Cross-Platform Performance and Power Prediction

Fast and accurate performance and power prediction is a key challenge in pre-silicon design evaluations during the early phases of hardware and software co-development. Performance evaluation using full-system simulation is prohibitively slow, especially with real world applications. By contrast, analytical models are not sufficiently accurate or still require target-specific execution statistics that may be slow or difficult to obtain. In this paper, we present LACross, a learning-based cross-platform prediction technique aimed at predicting the time-varying performance and power of a benchmark on a target platform using hardware counter statistics obtained while running natively on a host platform. We employ a fine-grained phase-based approach, where the learning algorithm synthesizes analytical proxy models that predict the performance and power of the workload in each program phase from performance statistics obtained on the host. Our learning approach relies on a one-time training phase using a target reference model or real hardware. We train our models on less than 160 programs from the ACM ICPC database, and demonstrate prediction accuracy and speed on 35 programs from SPEC CPU2006, MiBench and SD-VBS benchmark suites. Results show that with careful choice of phase granularity, we can achieve on average over 97% performance and power prediction accuracy at simulation speeds of over 500 MIPS.