PhD Thesis Defense: Nicolas Van Kempen, Improving Performance and Energy Efficiency with Native Languages and AI-Enabled Tooling
Content
Speaker:
Abstract:
Software performance is an increasingly important area of research, driven by sustainability concerns and the diminishing growth rate of hardware performance. This dissertation investigates whether programmers can improve the energy efficiency of their software, and how they can use compiled native languages alongside modern tooling to achieve faster and greener programs.
This dissertation first establishes and empirically validates a causal model of the relationship between programming languages and energy consumption. Using controlled experiments and improved measurement methodology, this causal model challenges prior work and demonstrates that the choice of programming language implementation has no measurable effect on application energy consumption beyond execution time. Optimizing software by choosing fast, compiled native languages and focusing on improving runtime performance will provide commensurate sustainability benefits.
This dissertation then identifies programmer-controlled memory management as a key mechanism enabling high performance in native languages. Memory is a primary bottleneck for performance; custom allocation strategies based on object lifetimes and memory access patterns can reduce allocation overhead and improve memory locality. Region-based custom allocators can improve performance by up to 15%. They are especially effective in latency-sensitive, long-running workloads: their performance stays predictable regardless of heap state, while naïve allocation can increase program execution time by as much as 2x with the default Linux allocator.
This dissertation finally introduces ChatDBG, an agentic debugging assistant that diagnoses low-level faults in native code. Aggressively optimizing at the memory level introduces memory management complexity that is a common source of failures. Effective performance work requires precise tooling to surface and resolve those issues. ChatDBG suggests an actionable fix for memory errors in real systems in 91% of runs, addressing the root cause in 36%.
Advisor:
Emery Berger