I/O Server Scheduling

This research involves the investigation of flow control (scheduling) techniques for multiple, I/O intensive parallel applications. Currently many of the parallel filesystems and I/O runtime libraries address maximizing the performance of individual parallel applications having intense I/O requirements. Most studies have not investigated the potential performance problems of handling tens to possibly hundreds of outstanding I/O requests in a multi-workload environment. Through some simple experiments, it has been found that disk scheduling may not perform well for a large number of outstanding requests for contiguous data on multiple disks. On single disk I/O servers, as I/O queues are filled with large requests spanning many files, there is the potential for high seek penalties that can lead to variability in request time, lower I/O performance and throughput.

The goals of this research will be to 1) develop some simple server scheduling policies and 2) investigate the benefits of these policies through performance evaluation experiments. The evaluation will be done using both simulation and a "straw-man" parallel filesystem. While the latter would allow experiments to be conducted on real machines, the availability and accessibility of very large systems would be a major concern as would be the amount of control over the variability in the system itself. At the cost of larger turn-around times on experiments, simulation allows very large systems to be simulated and detailed statistics to be gathered. Also, for system scaling studies, it is impossible to change the speeds of components (i.e., disk, cpu, network) in real systems. This is very easy to do in a simulator. The experiments will be driven by I/O traces from applications taken mostly from the scientific processing domain and also some from the non-scientific domain, including web-servers and interactive applications. The traces are gathered from applications ported and developed within the High Performance Software Laboratory at University of Maryland and also from the Scalable I/O Project (Specifically, The Pablo Group at UIUC).

Some of the expected benefits from this research will be 1) a better understanding of the behavior of parallel filesystems that are shared by multiple I/O intensive applications and 2) some flow control techniques that may improve server efficiency, application and system throughput.