ML for ML: Learning Cost Semantics by Experiment

Abstract

It is an open problem in static resource bound analysis to connect high-level resource bounds with the actual execution time and memory usage of compiled machine code. This paper proposes to use machine learning to derive a cost model for a high-level source language that approximates the execution cost of compiled programs on a specific hardware platform. The proposed technique starts by fixing a cost semantics for the source language in which certain constants are unknown. To learn the constants for a specific hardware, a machine learning algorithm measures the resource cost of a set of training programs and compares the cost with the prediction of the cost semantics. The quality of the learned cost model is evaluated by comparing the model with the measured cost on a set of independent control programs. The technique has been implemented for a subset of OCaml using Inria’s OCaml compiler on an Intel x86-64 and ARM 64-bit v8-A platform. The considered resources in the implementation are heap allocations and execution time. The training programs are deliberately simple, handwritten micro benchmarks and the control programs are retrieved from the standard library, an OCaml online tutorial, and local OCaml projects. Different machine learning techniques are applied, including (weighted) linear regression and (weighted) robust regression. To model the execution time of programs with garbage collection (GC), the system combines models for memory allocations and executions without GC, which are derived first. Experiments indicate that the derived cost semantics for the number of heap allocations on both hardware platforms is accurate. The error of the cost semantics on the control programs for the x86-64 architecture for execution time with and without GC is about 19.80% and 13.04%, respectively. The derived cost semantics are combined with RAML, a state-of-the-art system for automatically deriving resource bounds for OCaml programs. Using these semantics, RAML is for the first time able to make predictions about the actual worst-case execution time.