Click here for full text:
Efficient and Adaptive Proportional Share I/O Scheduling
Gulati, Ajay; Merchant, Arif; Uysal, Mustafa; Varman, Peter J.
Keyword(s): Storage QoS, scheduling, fair queueing, fairness, efficiency, storage resource sharing
Abstract: Quality of service (QoS) guarantees for applications are desirable under many scenarios. Despite much prior research on providing QoS in storage systems, current storage systems do not support extensive QoS guarantees. We believe this is mainly due to the low I/O efficiency of the various mechanisms designed for QoS. We find that I/O efficiency has received surprisingly little attention in storage QoS research. This is puzzling since the well known characteristics of I/O devices indicate that their efficiency depends crucially on the order in which the requests are served. In this paper, we attempt to alleviate the I/O efficiency concerns of proportional share schedulers. We first study the inherent trade-off between fairness and I/O efficiency. We find that significantly higher I/O efficiency can be achieved by slightly relaxing short-term fairness guarantees. We then develop several low-level mechanisms for proportional share schedulers and present a self-tuning algorithm that achieves good efficiency while still providing fairness guarantees. Experimental results indicate that an I/O efficiency of over 90% is achievable by allowing the scheduler to deviate from proportional service for a few seconds at a time.
Back to Index