A Taxonomy and Empirical Evaluation of WebAssembly Runtimes for Performance and Energy Across Machine Learning and Arithmetic Workloads

WebAssembly (Wasm) has emerged as a lightweight, portable, and secure execution environment beyond the browser, with growing adoption in data-intensive, machine learning, and performance-sensitive workloads. However, the runtime ecosystem remains fragmented, and a systems-level understanding of how runtime design affects performance and energy efficiency is still lacking. This paper makes two contributions. First, it presents a taxonomy of modern WebAssembly runtimes that classifies the runtime landscape by execution model, compilation strategy, integration style, and systems-level capabilities. Second, it provides an empirical evaluation of 13 WASI-compatible runtimes and a native baseline across eight benchmarks, namely, four machine learning workloads (Logistic Regression, K-Means, Naïve Bayes, and Decision Tree) and four arithmetic kernels (Matrix Multiplication, 2MM, Correlation, and Covariance) at three dataset scales. Execution time and dynamic energy are measured using an external hardware power meter under a statistically controlled repeated-run methodology. The results show strong workload dependence: optimized engines such as Wasmtime AOT, WasmEdge AOT, Wasmer LLVM, and WAMR AOT consistently achieve performance competitive with the portable native baseline. For regular arithmetic kernels, several optimized AOT runtimes match or slightly exceed the default Rust release build; however, these observations should be interpreted relative to a portable native reference configuration rather than a processor-specific optimized native implementation. Interpreter-oriented runtimes incur overheads exceeding 10× in both time and energy on iterative and branch-intensive workloads. Predictive-quality metrics remain effectively stable across runtime configurations, with only negligible variation, confirming that runtime choice primarily affects performance and energy rather than predictive quality. The findings establish that efficiency is determined primarily by runtime design rather than the WebAssembly execution model itself, and provide practical guidance for runtime selection in performance- and energy-sensitive computing contexts.