Falcon: A Scalable Analytical Cache Model
summary
Falcon builds a scalable analytical model for LRU cache miss rates using polyhedral dependence analysis and by splitting analysis to avoid blowup; it evaluates 46 TorchVision CNNs with geomean runtime 44.9s versus over 32 minutes for the prior model (the prior hit a four‑hour timeout on 54% of networks, Falcon never does, max 3m03s). The implementation parallelizes (16 cores → 8.2× speedup), fixes cache‑line padding approximations, and supports 513 ms incremental updates with measured prediction correlation ~0.98 versus ~0.56 for prior work.
AI-written summary · about
chapters (16)
- 0:00Falcon cache model intro
- 1:49How LRU caches work
- 3:38State-of-the-art model limits
- 4:21Dependence analysis for cache
- 6:52Optimizing dependence analysis
- 7:36Incremental updates in Falcon
- 9:41Impact of incremental analysis
- 10:02Experimental setup
- 11:07Performance vs HaStar
- 12:12Accuracy vs HaStar
- 13:14Cache lines and alignment
- 14:38Q&A: Cache property inheritance
- 15:20Q&A: Unknown trip counts
- 15:41Q&A: Multi-threaded cache pollution
- 16:05Q&A: Hardware prefetching impact
- 16:47Q&A: Compiler optimization interaction
related talks
01h 23mLoad balance and cache affinity: Friends, foes, or frenemies?
12mCascade models in computer vision to boost accuracy and performance
40m"Introduction to Face Processing with Computer Vision|"
17mRECL: Responsive Resource-Efficient Continuous Learning for Video Analytics
15mBoggart: Towards General-Purpose Acceleration of Retrospective Video Analytics
15mVideo Analytics with Zero-streaming Cameras
17mVision AI in 2025
52mFace Recognition: From Scratch To Hatch