Co-Design for High-Performance Computing Software Systems
summary
Co-design work shows concrete wins and concrete pain points: a 4M-line quantum chemistry code spends 15% idle time waiting on emulated one-sided RMA, so adding software-driven progress threads (extra CPU cores) recovers performance to near hardware-handled levels; MPI + OpenMP hybrid apps suffer from single-threaded communication and MPI global locks that block concurrency; profiling-based autoscaling lags application-guided tuning by ~10%; custom low-level thread/runtime changes yield up to 4× speedup on hotspots.
AI-written summary · about
chapters (11)
- 0:00Start
- 3:39Trend of HPC Architecture
- 8:25Deep HPC Stack: NWChem
- 12:33Performance Challenges in NWChem Deep Stack
- 20:41Performance Challenges in QMCPACK Wide Stack Lack of interoperability between runtime libraries
- 26:56Co-Design for Communication Resource Management Application - Communication Runtime Co Design
- 30:53NWChem with Adaptable Communication Resource Management
- 34:23Take Away from Communication Resource Management
- 35:41Co-Design for MPI and Threads Runtime Coordination
- 37:16Co-Design for Low-Overhead Multithreading Communication
- 40:35Challenges and Opportunities in Co-Design for HPC Software • HPC is powerful but brings one of the most complex software/hardware stack in
related talks
51mDesigning for Performance
23mMake Quantum Computing Useful with Performance Enhancing Infrastructure Software
21mSoftware for Quantum Computing- Easing Solutions for Complex
13mCohort: Software-Oriented Acceleration for You, Me, and Our Heterogeneous SoCs
01h 01mProgramming Abstractions for Orchestration of HPC Scientific Computing
01h 04mThe Future of Fast Code: Giving Hardware What It Wants
30mCINECA OpenPower-based HPC Infrastructure: Some porting results
12mAccelerating Software Development for High Performance Chiplet...