Benchmarking Graphical Models on POWER with Healthcare...
summary
Parallelized hill-climbing Bayesian network structure learning using pgmpy, Dask APIs and a Hadoop wrapper runs on an HPC POWER9 IC922 (160 vcores) against an HPE ProLiant DL580 Intel Xeon node; experiments use a 10.5M-record HEPMUS dataset (28 variables, binarized) and compare BIC and K2 scores. POWER9 shows 105× single-thread speedup in some runs, a consistent ~1.5× advantage in multicore tests, linear scaling with SMT, and statistically significant results after distributing tasks beyond 32 map tasks.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 0:45What are Graphical Models
- 1:43Why Graphical Models
- 4:22Structural Learning
- 5:53Types of Data
- 6:58Crosssectional Data
- 7:43Exhaustive Enumeration
- 8:09Marker Equivalence
- 8:45Health Clamping
- 9:32Hill Climbing Approach
- 12:52Implementation
- 15:06Performance Results
- 17:52Medicine Nexus
related talks
42mDifferential Software Solutions on HPC and IBM Power Architecture
43mAI APPLICATIONS ON POWER9 SYSTEM
45mHigh-Performance and Scalable Middleware for HPC and Deep...
44mArchitecture Effects When Interacting with GPU Technologies
31mHigh-Performance MPI and Deep Learning on OpenPOWER Platform
15mScaling the Next Paradigm of Heterogeneous Intelligence
01h 06mBodo: Automatic HPC Performance and Scaling for Data Processing in Python
13mFrom Silicon to Scale: Training at Scale