High-performance computing (HPC) systems play an essential role in large-scale scientific computations. As the number of nodes in HPC systems continues to increase, their power consumption leads to larger energy costs. The energy costs pose a financial burden on maintaining HPC systems, which will be more challenging on future extreme-scale systems where the number of nodes and power consumption are expected to further grow. To support this growth, higher degrees of network and memory resource sharing are implemented, causing a substantial increase in performance variation and degradation. These challenges call for innovations in HPC system middleware that reduce energy cost without trading off performance.
By taking the performance of an HPC system as a first-order constraint, this thesis establishes that HPC systems can participate in demand response programs while providing performance guarantees through a novel design of the middleware. Well-designed middleware also enables enhanced performance by mitigating resource contention induced by energy or cost restrictions. This thesis aims to realize these goals through two complementary approaches.
First, this thesis proposes novel policies for HPC systems to enable their participation in emerging power markets, where participants reduce their energy costs by following market requirements. Our policies guarantee that the Quality-of-Service (QoS) of jobs does not drop below given constraints and systematically optimize cost reduction based on large deviation analysis in queueing theory. Through experiments on a real-world cluster whose power consumption is regulated to follow a dynamically changing power target, this thesis claims that HPC systems can participate in emerging power programs without violating the QoS constraints of jobs.
Second, this thesis proposes novel resource management strategies to improve the performance of HPC systems. Better resource management can mitigate contention that causes performance degradation and poor system utilization. To resolve network contention, we design an intelligent job allocation policy for HPC systems that incorporate the state-of-the-art dragonfly network topology. Our allocation policy mitigates network contention, reduces network communication latency, and consequently improves the performance of the systems. As some latest HPC systems support the collection of high-granularity network performance metrics at runtime, we also propose a method to quantify the impact of network congestion and demonstrate that a network-data-driven job allocation policy improves HPC performance by avoiding network traffic hot spots. / 2022-01-18T00:00:00Z
| Identifer | oai:union.ndltd.org:bu.edu/oai:open.bu.edu:2144/41893 |
| Date | 19 January 2021 |
| Creators | Zhang, Yijia |
| Contributors | Coskun, Ayse K. |
| Source Sets | Boston University |
| Language | en_US |
| Detected Language | English |
| Type | Thesis/Dissertation |
Page generated in 0.0023 seconds