TY - GEN
T1 - Analyzing users in parallel computing
T2 - 14th International Conference on High Performance Computing and Simulation, HPCS 2016
AU - Schlagkamp, Stephan
AU - Ferreira da Silva, Rafael
AU - Renker, Johanna
AU - Rinkenauer, Gerhard
N1 - Publisher Copyright:
© 2016 IEEE.
PY - 2016/9/13
Y1 - 2016/9/13
N2 - The performance evaluation of parallel computing environments is crucial for the design of parallel job schedulers, as well as policy definitions. The analysis of user behavior is fundamental to unveil individual behaviors and reactions to different system performances (e.g., scarce resources, low throughput, etc.). In this paper, we present an analysis of parallel computing users based on responses to the Questionnaire for User Habits of Computer Clusters (QUHCC). The survey is composed of 7 measures and 53 items, and was answered by 23 users of computer clusters at TU Dortmund University. We investigate several influences on working behavior, including the influence of slow responses on working times, strategies to cope with high contention and poor performance, user's experience, and user satisfaction. Analysis results reveal that user satisfaction is negatively correlated to the application slowdown; users tend to work after hours to improve their efficiency; informal agreements between users are established to coordinate executions and reduce the system load; and scientific experiments may include several clusters, thus the user submission behavior should be seen from a multi-dimensional perspective. We then compare and discuss the analysis results with conclusions obtained from statistical trace analysis to reveal unknown and hidden correlations and feedbacks between system characteristics and the subsequent job submissions. Our findings indicate that the user characteristics together with the historical information (traces) are crucial to build a concise understanding of feedback effects between the user satisfaction, their job submission behavior, and the system performance. Additionally, this paper also provides a first overview of which user reactions may be the most relevant for dynamic performance evaluation.
AB - The performance evaluation of parallel computing environments is crucial for the design of parallel job schedulers, as well as policy definitions. The analysis of user behavior is fundamental to unveil individual behaviors and reactions to different system performances (e.g., scarce resources, low throughput, etc.). In this paper, we present an analysis of parallel computing users based on responses to the Questionnaire for User Habits of Computer Clusters (QUHCC). The survey is composed of 7 measures and 53 items, and was answered by 23 users of computer clusters at TU Dortmund University. We investigate several influences on working behavior, including the influence of slow responses on working times, strategies to cope with high contention and poor performance, user's experience, and user satisfaction. Analysis results reveal that user satisfaction is negatively correlated to the application slowdown; users tend to work after hours to improve their efficiency; informal agreements between users are established to coordinate executions and reduce the system load; and scientific experiments may include several clusters, thus the user submission behavior should be seen from a multi-dimensional perspective. We then compare and discuss the analysis results with conclusions obtained from statistical trace analysis to reveal unknown and hidden correlations and feedbacks between system characteristics and the subsequent job submissions. Our findings indicate that the user characteristics together with the historical information (traces) are crucial to build a concise understanding of feedback effects between the user satisfaction, their job submission behavior, and the system performance. Additionally, this paper also provides a first overview of which user reactions may be the most relevant for dynamic performance evaluation.
KW - User behavior analysis
KW - parallel computing
KW - user satisfaction
UR - http://www.scopus.com/inward/record.url?scp=84991608308&partnerID=8YFLogxK
U2 - 10.1109/HPCSim.2016.7568362
DO - 10.1109/HPCSim.2016.7568362
M3 - Conference contribution
AN - SCOPUS:84991608308
T3 - 2016 International Conference on High Performance Computing and Simulation, HPCS 2016
SP - 395
EP - 402
BT - 2016 International Conference on High Performance Computing and Simulation, HPCS 2016
A2 - Zeljkovic, Vesna
A2 - Smari, Waleed W.
PB - Institute of Electrical and Electronics Engineers Inc.
Y2 - 18 July 2016 through 22 July 2016
ER -