Quantifying System-Level Optimizations for Latency and Frame-Time Stability in Real-Time Workloads
Resumo
This paper investigates how Windows 11 and GPU driver tuning affect latency-related behavior in a real-time gaming workload. We compare an out-of-the-box profile (OOB) and an optimized profile (OPT) in Counter-Strike 2 (CS2) using PresentMon, FrameView, VProf Lite, and LatencyMon. Across ten deterministic low-preset runs per profile, OPT increases average throughput from 272.64 to 281.61 FPS (+3.29%), but worsens frametime tails (p95: 5.520 to 6.987 ms; p99: 6.017 to 7.795 ms). A complementary preset comparison shows that graphics-load choice has a much larger effect than OS tuning: the high preset drops performance to about 132 FPS with 7.56 ms mean frametime. Overall, the results expose a clear trade-off between average throughput and worst-frame behavior and reinforce the use of tail-aware metrics in latency-sensitive workloads.
Referências
Montgomery, J. and Intel (2023). Presentmon: Capture and analysis tool for frame timing and latency on windows. GitHub / Intel / Open-source tool.
NVIDIA (2020). Frameview user guide: Performance, latency and power analysis tool. NVIDIA Documentation.
Spjut, J., Boudaoud, B., and Kim, J. (2021). A case study of first person aiming at low latency for esports. NVIDIA Research.
Terhell, D. (2014). Windows and real-time.
Terhell, D. (2023). Latencymon technical overview.
Tokey, S. S., Boudaoud, B., Kim, J., Spjut, J., and Claypool, M. (2025). Timing matters: The impact of event-specific frametime spikes in first-person shooter games. In 2025 17th International Conference on Quality of Multimedia Experience (QoMEX), pages 1–7.
