Scheduling for improved write performance in a Cost-Effective, Fault-Tolerant 
Parallel Virtual File System (CEFT-PVFS)

Yifeng Zhu, Hong Jiang, Xiao Qin, Dan Feng and David R. Swanson

Department of Computer Science and Engineering
University of Nebraska-Lincoln
Lincoln, NE 68588-0115, {yzhu, jiang, xqin, dswanson}@cse.unl.edu 

Without any additional hardware, CEFT-PVFS utilizes the existing disks on each cluster node to provide RAID-10 style parallel I/O service. In CEFT-PVFS, all servers are also computational nodes and can be heavily loaded by different applications running on the cluster, thus potentially degrading the I/O performance. To minimize the degradation, I/O requests can be scheduled on a less loaded server in each mirroring pair. To help define the meaning of “load” in face of multiple resources such as CPU, memory, disk and network, this paper examines the impacts of these resources by measuring aggregate I/O throughput of the simplest CEFT-PVFS configurations, under specific and isolated workload stresses. Based on the heuristic rules found from the experimental results, a scheduling algorithm for dynamic load balancing is developed. In a CEFF-PVFS with 16 data servers, we evaluate this algorithm under different workloads. The results show that the proposed scheduling algorithm significantly improves the overall performance.

ClusterWorld Conference and Expo Partners with the Fourth LCI International Conference on Linux Clusters:
The HPC Revolution 2003, San Jose, California, June 24-26, 2003. 

---