Automatically Scalable Computation
summary
Claims a runtime that learns program state trajectories and speculatively executes predicted future states, using a software x86 emulator (scaly ASK) and a native Pin/ptrace-based system (speedy ASK). Predictors include last-value, mean, logistic and 32-bit linear models; dependency tracking builds read/write masks into a trajectory cache. Results show surprising wins: memoization speeds some workloads on one core (Collatz) and near-linear scaling to ~100 cores and useful speedup up to ~2,000 cores before linked-list limits.
AI-written summary · about
chapters (30)
- 0:00Intro
- 0:20Automatically Scalable Computation
- 1:57A Different Approach
- 2:26Could Computers Get Faster With Practice?
- 4:02Computers can get Faster!
- 4:50Join me in a thought experiment...
- 5:16Execution in a Really Big State Space
- 10:52Some intuition
- 11:40An Automatically Scalable Architecture
- 14:26Two Implementations
- 15:04Implementation: ScaleyASC
- 16:17Scaley ASC: VM
- 16:47Scaley ASC: Recognizer
- 18:02Scaley ASC: Predictors
- 20:06Scaley ASC: Allocator
- 21:43Scaley ASC: Speculators
- 24:41Dependency Tracking- Symmetries
- 25:18Scaley ASC: Trajectory Cache
- 26:22How Well Does it work?
- 27:41Even Cooler Results
- 30:01Another Example
- 30:37Implementation: Speedy ASC
- 32:16SpeedyASC: Critical Technology
- 33:41SpeedyASC: The Recognizer
- 34:53SpeedyASC: The Monitor
- 36:30Revisiting our Pintool (dependency tracking)
- 37:19SpeedyASC: The Trajectory Cache
- 41:01The Allocator
- 45:55Recall: The Monitor
- 47:13Where to Next?
related talks
11mAccelerated Software for a Post-Moore World
01h 03mScaling up machine learning without tears (and what do programming languages ha...
47mMegacore: Megafast, Megacool?
01h 02mScheme as a framework for Deep Learning
36mAccelerating AI training fleets with sched_ext
32mProbabilistic and Differentiable Programming in Scientific Simulators
35mScalable structure learning and inference for domain-specific probabilistic prog
11mEngineering Scalable ML Infrastructure for Modern Apps