A Methodology for Oracle Selection of Monitors and Knobs for Configuring an HPC System running a Flood Management Application

This paper defines a methodology for the oracle selection of the monitors and knobs to use to configure an HPC system running a scientific application while satisfying the application’s requirements and not violating any system constraints. This methodology relies on a heuristic correlation analysis between requirements, monitors and knobs to determine the minimum subset of monitors to observe and knobs to explore, to determine the optimal system configuration for the HPC application. At the end of this analysis, we reduce an 11-dimensional space to a 3-dimensional space for monitors and a 6-dimensional space to a 3-dimensional space for knobs. This reduction shows the potential and highlights the need for a realistic methodology to help identify such minimum set of monitors and knobs.

💡 Research Summary

The paper addresses the challenge of configuring high‑performance computing (HPC) systems for a specific scientific workload while satisfying multiple quality‑of‑service (QoS) constraints such as performance, power, energy, availability, and cost. The authors propose an “oracle selection” methodology that identifies the minimal set of system monitors and configurable knobs (hardware/software parameters) required to achieve an optimal configuration for the target application.
The case study is Floreon+, an online flood‑prediction and management application that, in its emergency mode, must complete 20 000 Monte‑Carlo iterations within ten minutes, maintain an availability of at least 0.99, keep power consumption below 81 W, and limit total energy to 48 600 J. To meet these constraints, the authors initially consider eleven monitors (execution time, IPC, MPKI, CPU and DRAM power, CPU temperature, server and system MTBF, etc.) and six knobs (DVFS frequency levels, Simultaneous Multi‑Threading, DRAM protection mode, Turbo mode, hardware prefetchers, and core redundancy).
Experiments are performed on a dual‑socket Intel Xeon E5‑2640 v3 cluster. All 128 possible combinations of the six knobs (each with two or more settings) are executed five times; the eleven monitor values are recorded, and the five QoS metrics are derived (performance from execution time, power from CPU/DRAM measurements, energy from power × time, availability from analytical MTBF models, and cost from CAPEX/OPEX estimation tools). The resulting dataset is a 128 × 11 × 5 tensor (knob configurations × monitor readings × requirements).
The core of the methodology is a heuristic Pearson‑correlation analysis. First, pairwise correlations among the five requirements are computed; any requirement that is highly correlated (≥ 0.9) with another and has the lowest average correlation with the rest is removed. The same procedure is applied to the monitor set and then to the knob set, each time retaining only those elements that contribute the most independent information with respect to the remaining requirements. After iterative pruning, the dimensionality is reduced from eleven monitors to three (execution time, CPU power, CPU temperature) and from six knobs to three (DVFS, SMT, DRAM protection).
The reduced set still satisfies the majority of the 128 configurations (over 85 % meet all QoS constraints), demonstrating that many monitors and knobs are redundant for this workload. The analysis also uncovers non‑obvious relationships, such as a strong positive correlation between DRAM power and CPU temperature, and a significant impact of SMT enablement on energy efficiency, which would be difficult to detect without systematic correlation study.
The authors acknowledge that the current approach requires exhaustive exploration of the full knob space to compute reliable correlations, which may be prohibitive for larger systems. They suggest future work to integrate sampling‑based meta‑models, machine‑learning predictors, or Bayesian optimization to reduce the initial data‑collection cost, and to validate the methodology on a broader range of HPC applications.
In conclusion, the paper demonstrates that a statistically‑driven reduction of monitors and knobs can dramatically shrink the configuration search space— from an 11‑dimensional monitor space to 3 dimensions and a 6‑dimensional knob space to 3 dimensions—while still meeting stringent application requirements. This provides a practical foundation for automated, QoS‑aware HPC system tuning and highlights the importance of data‑driven methodologies in large‑scale system configuration.