The Death of Thread per Core
2 days ago
- #work-stealing
- #data-processing
- #async-runtimes
- Async runtimes handle long-running tasks that can yield and spawn new work.
- Work-stealing models allow threads to balance work by moving tasks between queues.
- Thread-per-core models focus on data locality but may suffer from work imbalance.
- Data processing initially favored thread-per-core for cache efficiency and simplicity.
- Recent trends suggest dynamic reshuffling in data processing can handle skew better.
- Increased core counts and improved IO latency make work-stealing more attractive.
- Shared-state concurrency is popular due to better scheduler insights into work types.
- Cultural shifts towards elasticity in systems to handle unpredictable skew at scale.