Self-Organizing Middleware for Extreme Heterogeneity: The Role of Technology- Oblivious Machine Learning

Extreme Heterogeneity (EH) resulting from diverse processors, accelerators, and memory devices impose urgent challenges to support scientific workflows in the post-Moore era. The position taken is that in-the-loop runtime machine learning and big-data approaches can offer feasible and efficient mechanisms to manage vast resources by realizing a thin autonomous layer of control and optimization. Research on autonomous middleware, residing between diverse hardware resources and existing system software primitives, is worthy of community discussion towards developer-transparent support for EH challenges. Furthermore, by prioritizing a self-aware, selforganizing, and self-optimization perspective, a thin middleware layer could be readily-transportable despite changes to the underlying computing stack. Overall, it will be advocated such an approach could transform EH from a “formidable foe of HPC” into a “valuable ally of HPC” on an incremental path to 2025-2040. Keywords—Runtime Systems for Extreme Heterogeneity; Machine Learning; Data Clustering; Fault-Propagation Classifier; Correctness Despite Silent Data Corruption; EnergyAware High Performance Computing; Runtime Dynamic Resource Optimization. I. CHALLENGES & PERSPECTIVE OF EXTREME HETEROGENEITY This position paper advocates the potential of online and offline machine-learning approaches to provide a feasible yet high-payoff computer science research direction to address the following Extreme Heterogeneity (EH) Challenges: C1: efficient resource utilization across disparate computing paradigms/components including multi/manycore CPU, GPU, TPU, FPGA, and future neuromorphicbased accelerators, C2: mitigating the perils of Silent Data Corruption (SDC) propagation that are not neglectable in HPC applications at the vastness of post-exascale component cardinalities, and C3: intractability for human optimization of resilience vs. energy tradeoffs [4], whereby reducing supply voltage or cooling to save energy may actually increase energy consumption, because running time is lengthened due to errors impacting convergence. Research Direction: Unsupervised and semi-supervised learning offer promising research approaches to EH challenges. Large-scale analysis of operational measures and vulnerability metrics could provide a wealth of orthogonal input data to machine learning approaches. Aided by its diversity, such data empowers runtime and offline learning strategies developed to attain preestablished {throughput, resilience, energy} objectives [6][8][11]. These could be researched for the execution behavior of individual jobs first and then layered later for a typical multiuser scenario. State of the Art: Energy-aware HPC methods have an extensive basis, as do various methods for achieving sufficient correctness despite SDC during execution. However, the management of both of these goals simultaneously as competing objectives remains an urgent, vital, and open research issue. Moreover, EH will act to further exacerbate their interaction and intractability to manage without some form of machine learning support. It is promulgated herein that feasible extensions to existing infrastructures for runtime dynamic resource