On Task Mapping Algorithm For Co-optimizing Computation And Communication Performance In Networks-on-chip

As computation demand of applications continues to grow,multi-/many-core chips are widely used in systems.In task mapping,each task of an application is mapped to the core in multi-/many-core chips.Task mapping has an impact on thermal distribution of chips.For example,mapping an application with high computation demand to a contiguous region might leads to thermal hotspots.Task mapping also determines the communication distance of cores.If two tasks are mapped to the cores that are far away from each other,the application communication latency is high.Therefore,task mapping is key to system performance.On a many-core chip,through power-gating a free core,the neighboring cores can have better heat dissipation,thus can run with a higher frequency.These power-gated cores are called dark cores.When an application is mapped to active cores,a few dark cores can also be allocated to the same application,so that it can run with high performance.However,including dark cores into region has disadvantages,such as increasing the communication distance of cores.When designing a task mapping algorithm,there are six crucial factors that must be considered.They are 1)the number of dark cores assigned to each application,2)the location of the dark cores,3)the features of applications,4)the features of Networks-on-Chip(No Cs),5)system workload,and 6)the task-to-core mapping.So far,the existing task mapping algorithms do not consider all these important factors,which influences system performance.In this paper,first,a performance model is proposed to predict the throughput of an application,according to the characteristics of applications and the number of dark cores assigned to it.A task mapping problem is formulated to maximize the overall system throughput,under the temperature constraint.For 2D No Cs,a task mapping algorithm is proposed,which first determines the number of dark cores assigned to each application,and allocates a core region for each application.After that,it maps each task to a core,and determines which cores are dark cores.In experiments,compared with other algorithms,the proposed method for 2D No Cs achieves 62% overall system performance improvement.Compared with 2D No Cs,the heat dissipation ability of cores in each layer of a 3D No C is different.The number of a region's layer influences communication latency.For 3D No Cs,we propose a task mapping algorithm which takes the locations of dark cores,the number of dark cores,the features of applications,system workload,the heat dissipation of cores,and the shape of application region into account.Compared with other algorithms,experimental results show that the proposed algorithm improves overall system performance by as much as 50%.