אירועים והרצאות בפקולטה למדעי המחשב ע"ש הנרי ומרילין טאוב

The diversity of workloads, performance metrics, and potential system optimizations makes designing an optimal caching middleware a daunting task. For example, cache replacement policies can exhibit widely varying hit ratios depending on workload characteristics and cache sizes. Moreover, the choice of cache replacement policy may impact the computational complexity and meta-data size required for cache management, which may also be impacted by the internal cache organization (e.g., fully associative vs. k-way) and supporting data structures. Consequently, state-of-the-art caching libraries are still hand-crafted by experts and carefully tuned to narrow workload classes and design goals. Adapting a cache to a new setting, or improving it as a robust general-purpose solution, still requires a labor-intensive redesign process that does not scale to the diversity of modern workloads and deployment constraints. We present Evolve, an agentic framework for automatically improving caching middleware implementations.

Starting from an existing design, Evolve searches for superior variants under two regimes: refinement over a fixed workload suite, and workload specialization against a deployment-specific trace distribution. Each cycle proposes a cohort of LLM-implemented variants, validates them through deterministic correctness gates, scores survivors on a multi-dimensional plugin-based fitness vector, retains a Pareto front under a held-out trace suite to control for overfitting.

The framework also stores LLM-generated insights in a knowledge ledger that supports cross-iteration learning and guides future exploration. As a case study, we apply Evolve on a Dash-inspired associative cache, run it across diverse traces in both regimes, and characterize the variants it produces along each chosen fitness dimension.