COMP 7500 Advanced Operating Systems Performance Evaluation Dr

COMP 7500 Advanced Operating Systems Performance Evaluation Dr. Xiao Qin Auburn University http: //www. eng. auburn. edu/~xqin@auburn. edu Spring, 2012

Performance Evaluation I/O-intensive Workload Conditions • The mean slowdowns increase with the I/O load • IOCM-RE is the best scheme 2

Performance Evaluation (cont. ) Memory-Intensive Workload Conditions • IOCM-RE and MEM outperform the CPU-based and NLB schemes considerably. 3

Remote Execution vs. Preemptive Migration I/O-intensive workload Traces have 30% parallel jobs. • Preemptive migration outperforms remote execution under I/O-intensive workload conditions. 4

Real I/O-Intensive Applications • IOCM-RE and IOCM-PM benefit all I/O intensive applications • IOCM-RE and IOCM-PM yield approximately identical performances. 5

Outline n n n Motivations A Disk-I/O-Aware Load Balancing Policy with Remote Execution A Disk-I/O-Aware Load Balancing Policy with Preemptive Migration Evaluation of the two Disk-I/O-Aware Policies Load Balancing for Heterogeneous Clusters Contributions and Conclusions 6

Disk Heterogeneity Level The weight of disk performance of node i: Wdiski is a ratio between its performance and that of the fastest disk in the cluster. The disk heterogeneity level: the standard deviation of disk performance weights. Formally, 7

Identify an overloaded node Node i’s I/O resource is considered overloaded, if the I/O load index load. IO(i) is higher than an I/O threshold, threshold. IO(i), Accumulative I/O load imposed by the running tasks Fraction of the total I/O processing power on node i. 8

Impact of Heterogeneity on the Performance of Load-balancing Policies 9

Impact of Heterogeneity 298 ->325 123 ->127 • Mean slowdowns increase as the system heterogeneity increases. • IOCM-RE is less sensitive to changes in disk heterogeneity level. 10

Outline n n n Motivations A Disk-I/O-Aware Load Balancing Policy with Remote Execution A Disk-I/O-Aware Load Balancing Policy with Preemptive Migration Evaluation of the two Disk-I/O-Aware Policies Load Balancing for Heterogeneous Clusters Contributions and Conclusions 11

Summary n Two disk-I/O-aware load balancing policies (Remote execution and remote I/O) n Preemptive job migrations n Consideration of the heterogeneity of Resources n Support parallel jobs n A feedback control mechanism n A communication-aware load-balancing policy 12

Conclusions A family of I/O-aware load balancing policies n To achieve high performance under a wide spectrum of workload conditions n Improve Buffer Utilization n Preemptive migration is better than remote execution for I/O-intensive applications n 13

Future Research n. A predicting tool n Combine space- and time-sharing strategies n Incorporate Coordinated scheduling n Optimize data movement 14