A Feedback Control Mechanism for Balancing I/O- and Memory-Intensive Applications on Clusters
Xiao Qin, Hong Jiang, Yifeng Zhu, David R. Swanson · 2005
Abstra t. One ommon assumption of existing models of load balan ing is that the weights of resour es and I/O buer size are stati ally ongured and annot be adjusted based on a dynami workload. Though the stati onguration of these parameters performs well in a luster where the workload an be modeled and predi ted, its performan e is poor in dynami systems in whi h the workload is unknown. In this paper, a new feedba k ontrol me hanism is proposed to improve overall performan e of a luster with a general and pra ti al workload in luding I/O-intensive and memory-intensive load. This me hanism is also shown to be ee tive in omplementing and enhan ing the performan e of a number of existing dynami load-balan ing s hemes. To apture the urrent and past workload hara teristi s, the primary obje tives of the feedba k me hanism are: (1) dynami ally adjusting the resour e weights, whi h indi ate the signi an e of the resour es, and (2) minimizing the number of page faults for memory-intensive jobs while in reasing the utilization of the I/O buers for I/O-intensive jobs by manipulating the I/O buer size. Results from extensive tra e-driven simulation experiments show that ompared with a number of s hemes with xed resour e weights and buer sizes, the feedba k ontrol me hanism delivers a performan e improvement in terms of the mean slowdown by up to 282% (with an average of 125%). Key words. Feedba k ontrol, I/O-intensive appli ations, luster, load balan ing 1. Introdu tion. S heduling [16, 19 ℄ and load balan ing [1, 10 ℄ te hniques in parallel and distributed systems have been investigated to improve system performan e with respe t to throughput and/or individual