Model driven cache-aware scheduling of object oriented software for chip multiprocessors

Leveraging utilization of the shared caches of multicore processors is one of the heavily studied topics of today's chip multiprocessing community. Providing a scheduling mechanism that maximizes throughput by reducing miss-rates of shared caches and preserves the fairness of processor usage is in the center of this problem. Proposed scheduling algorithms in this field usually take advantage of thread level properties of software providing modifications at operating system level. In our study we choose to approach the problem from a different perspective and use software models to guide operating system to effectively map software's objects onto processor cores. In an object oriented software objects collaborate on fulfilling jobs and they may operate on common data. Our scheduling method takes class dependencies into account and tries to schedule objects of coupled classes onto cores that share the common cache. This paper presents case studies on implementations of three software design patterns(Strategy, Visitor and Observer) and an image filtering software implementation. During our experiments we use our cache-aware scheduler in guiding Linux's completely fair scheduler (CFS) to perform more cache-aware schedules and decrease running time around 10%. Our results promise that guiding/restricting operating system's scheduler using class-relational information present in the object oriented software model can be fruitful in increasing software performance on multicore processors.