Balancing performance and fairness using finite-SRPT

Abstract

Among all scheduling policies it has been well known that Shortest-Remaining-Processing-Time (SRPT) is optimal in mean delay. Previous reservations on SRPT due to possible unfairness has been shown to be have much less impact than feared. The conclusions point to the feasibility of SRPT to real web-servers to improve delay performance. However, SRPT is unrealistic when we consider web-servers with large bandwidth (BW) since files with the smallest remaining processing time are granted the full capacity of the server until it is preempted or is fully served. For web-servers with large BW, it is more realistic to serve multiple file requests by a fixed amount that corresponds to the link rates of the end-receivers. To address this problem, we propose finite-SRPT, a scheduling policy which ensures that all file requests are served a finite amount each time-slot and faithfully follows the basic operation of SRPT, i.e., files of smaller remaining processing time are served first. We investigate its delay distribution using a queueing model where the server is accessed by a large number of sources. When the maximum amount of service available to a file in a time-slot is larger than the largest file, finite-SRPT is equivalent to the original SRPT. As the maximum amount decreases, the finite-SRPT becomes similar to Processor-Sharing (PS). Thus finite-SRPT provides a way in which to balance delay performance and fairness by simply adjusting the maximum amount a file is served in a time-slot.