Cross-layer workload characterization of meta-tracing JIT VMs

Dynamic programming languages are becoming increasingly popular, and this motivates the need for just-in-time (JIT) compilation to close the productivity/performance gap. Unfortunately, developing custom JIT-optimizing virtual machines (VMs) requires significant effort. Recent work has shown the promiseofmeta-JITframeworks, which abstract the language definition from the VM internals. Meta-JITs can enable automatic generation of high-performance JIT-optimizing VMs from high-level language specifications. This paper provides a detailed workload characterization of meta-tracing JITs for two different dynamic programming languages: Python and Racket. We propose a new cross-layer methodology, and then we use this methodology to characterize a diverse selection of benchmarks at the application, framework, interpreter, JIT-intermediate-representation, and microarchitecture level. Our work is able to provide initial answers to important questions about meta-tracing JITs including the potential performance improvement over optimized interpreters, the source of various overheads, and the continued performance gap between JIT-compiled code and statically compiled languages.