High Speed CPU Simulation using JIT Binary Translation
- Type: Paper
- Authors:
N.P. Topham and D. Jones. - The 3rd Annual Workshop on Modeling, Benchmarking and Simulation, held in conjunction with ISCA-34, San Diego CA.
- Download as PDF
-
Abstract:
Instruction set simulators are indispensable tools for exploring the design-space of innovative processor architectures, for processor verification, and for software development. Traditional interpretive simulators are too slow to cope with the increasing complexity of embedded processors now being deployed in many high performance systems. High speed emulation techniques based on dynamic binary translation have been proposed previously, but thus far we have not seen flexible multi-function full-system simulators capable of acting as golden reference models, software development platforms and design-space exploration tools. This paper presents a target-adaptable full-system simulator which combines the speed of JIT binary translation with the observability of interpreted simulation. We explain the mechanisms it uses to achieve sufficiently high performance to boot and run Linux interactively at speeds exceeding those achievable with FPGA-based RTL emulation of the same processor. We report performance figures from a set of representative embedded benchmarks which range from 187 to 373 MIPS. Our results also indicate that transient simulation speeds can exceed 1,000 MIPS, and we show that a full-system Linux simulation can sustain more than 148 MIPS.
