Fixed Share Scheduling via Dynamic Weight Adjustment in Proportional Share Scheduling Systems

Abstract

There exist many compute-intensive softreal-time applications, such as video encoding/decoding and data encryption/decryption, that require a fixed percentage of CPU cycles to maintain an acceptable QoS (Quality of Service). Unfortunately, major operating systems do not support well such compute-intensive softreal-time applications, since they commonly rely on a proportional share policy. In this paper, we present a novel scheduling policy, called FSS (fixed share scheduling), to enable a proportional share scheduler to support compute-intensive soft real-time applications as well as non-real-time applications. The goal of FSS is to guarantee an absolute, constant share of CPU cycles for soft real-time tasks regardless of workload conditions, whereas traditional proportional share schedulers focus on relative proportional guarantees. To do so, working on top of a proportional share scheduler, FSS dynamically changes the weight value of each softreal-time task to match the demanded amount of CPU share under varying workload conditions. The weighted fairness mechanism of the underlying proportional share scheduler will then provide the demanded amount of CPU share for each softreal-time task. To demonstrate the efficacy of FSS, we have implemented a fixed shared scheduling prototype in the Linux CFS (completely fair scheduler) and conducted experiments to show the correctness and efficiency of the FSS scheme.