Towards Soft Optimization Techniques for Parallel Cognitive Applications

1. Towards Soft Optimization Techniques for Parallel Cognitive Applications Woongki Baek, JaeWoong Chung, Chi Cao Minh Christos Kozyrakis, Kunle Olukotun Computer Systems Laboratory Stanford University http://tcc.stanford.edu

2. Introduction • Cognitive applications are widely used • From commercial fields to military and security applications • Efficient parallelization techniques are needed • To process ever increasing data sets • This brief announcementdiscusses: • Optimizations that build upon soft computing properties • How they target common bottlenecks for parallel execution • Case study with Loopy Belief Propagation (LBP)

3. Soft Computing Properties • Conventional vs. cognitive applications • Conventional: precise data, strict accuracy requirements • Cognitive: inherently noisy inputs, must handle uncertainty, acceptable approximation of the “correct” answer • Soft computing properties: • User-defined, relaxed correctness • E.g., a small percentage of misclassification rate is tolerable • Redundancy • E.g., computations and communications on converged nodes • Inherent adaptivity to errors • E.g., noisy data from sensor nodes in online object tracking

4. Soft Optimizations (1) • O1) Reducing computation • Target: excessive workload due to large data sets • Possible optimizations: • Data dropping (sampling) • Lazy computation • Aggressive solution pruning • O2) Mitigating imbalance • Target: work and region imbalances • Possible optimizations: • Adaptive workload discarding • Selective barriers

5. Soft Optimizations (2) • O3) Reducing communication • Target: expensive communications on large-scale systems • Possible optimization: • Adaptive communication • O4) Reducing synchronization • Target: reduce synchronization frequency • Possible optimizations: • Selective synchronization • Imprecise updates

6. Case Study: Loopy Belief Propagation (LBP) • Efficient approximation for probabilistic inference on graphical models • Computes & communicates “beliefs” of nodes in graph. • Complexity grows linearly with number of nodes. • Not guaranteed to give the correct beliefs for networks with loops.

7. Evaluation • Soft optimizations on LBP • Adaptive message version 1 (MSG1) • Discardsmessages when both a sender and a receiver have converged • Target: O1, O2, and O3 • Adaptive message version 2 (MSG2) • Discards messages when only a sender has converged • Target: O1, O2, and O3 • Lazy belief computation (LazyBC) • Skip belief computations on converged nodes • Target: O1 and O3 • Evaluation • A detailed simulator for a shared-memory multiprocessor • Metrics: performance improvement & accuracy loss

8. 2.8 2.7 3.7 4.8 2.4 Results: Speedup • #1: Soft optimizations improve performance significantly • E.g., 56.1x on 32 processors (superlinear) #2: Consistent benefits over conventional forall processor counts

9. Results: Accuracy • Outstanding tolerance to soft optimizations (accuracy loss < 0.2%) • Note: NOT a proof for general accuracy guarantees • Further research necessary on this question • Can use soft versions for real-time answers verified later on

10. Implications • We have shown that soft optimizations can be valuable • Further research • Algorithms • Evaluation with more applications • Analysis of convergence, accuracy, and performance • Programming languages and runtime systems • Language constructs to express soft optimizations • Runtime support for monitoring & automatic adaptation • Architectures • Support for relaxed coherence or communication • Support for introspection

11. Thanks & Questions? Woongki Baek (wkbaek@stanford.edu)

12. Backup: Simulated Architecture

13. Backup: Performance & Accuracy (32 CPUs)