Achieving top performance on modern many-core, accelerated, multi-node systems is a daunting challenge. Production software systems suffer from performance losses at all layers of the software stack, which can be broadly attributed to three main causes: resource idleness, wasteful resource consumption, and insufficient tailoring for architectural characteristics. Effective software tools, adaptive runtime systems, and efficient algorithms developed as part of my doctoral dissertation address instances of performance problems due to each of the aforementioned causes. Tools and techniques that I developed have demonstrated their effectiveness by identifying performance losses arising from developer’s inattention to performance, inappropriate choice of data structures, inefficient algorithms, use of heavyweight abstractions, and ineffective compiler optimizations. This work has had practical impact through collaborations with various national laboratories (LBNL, LLNL, ORNL, and PNNL), industrial, and academic partners. An adaptive runtime developed in collaboration with LBNL eliminated more than 60% redundant barriers in production runs of NWChem - a flagship DOE computational chemistry code. A fine-grained instruction monitoring framework pinpointed inefficiencies, which helped us substantially improve the performance of several important codes. Idleness analysis for heterogeneous architectures, developed as part of Rice University’s HPCToolkit performance tools, helped us diagnose and correct both hardware and software causes of performance losses for important codes running on accelerated supercomputers. Novel, architecture-aware synchronization algorithms deliver high throughput, high fairness, and superior scaling of highly contended locks on deep NUMA architectures.
