Reducing Cold-Start Effects in Hybrid Cycle-Accurate / Abstract CPU Simulation Using Accelerated Pre-Warmup 94 views

Cycle-accurate simulation is essential for modern CPU performance analysis but often suffers from cold-start effects, where caches and internal states begin in unrealistic empty conditions, leading to significant accuracy deviations and extended warm- up times. This thesis presents Accelerated Pre-Warmup, a scalable framework that reconstructs realistic cache and microarchitectural states prior to detailed simula- tion. The proposed method integrates per-thread prewarmup trace generation and fast-forward execution control within a hybrid simulation environment that com- bines cycle-accurate and abstract functional models. By capturing representative L2/L3 cache transactions and replaying them during initialization, the framework mitigates cold-start bias while maintaining simulation determinism. The implemen- tation extends the existing hybrid Detailed/abstract CPU infrastructure to support multi-device synchronization, configurable fast-forward counts, and state reuse across threads. Experimental results demonstrate that Accelerated Pre-Warmup reduces warm-up latency by up to 3.2× and improves cache-hit estimation accuracy within 1.8% of steady-state behavior, with negligible runtime overhead. The proposed tech- nique enables faster, more faithful performance characterization of large-scale CPU designs, offering a practical balance between simulation fidelity and efficiency.

MS (Master of Science)

All rights reserved by the author (no additional license for public reuse)

2025-12-10