ichat: Inter-cache hardware-assistant data transfer for heterogeneous chip multiprocessors

Modern heterogeneous multiprocessors integrate CPU and GPU together to provide a boost to computational perfor-mance. With tighter integration of CPU and GPU, it is criti-cal to share and move data more efficiently in order to lever-age the computational power that a GPU can provide. Ini-tially, DMA or PCIe devices were used to transfer data be-tween CPU and GPU with low efficiency and little flexibil-ity. Recently single address space and coherent cache hier-archies are being adopted in heterogeneous architectures to share data more efficiently. Thus it is becoming critical to understand the communications overheads in this new con-text and to improve communication efficiencies for these ar-chitectures. This paper proposes a novel approach called iCHAT (inter-Cache Hardware-Assistant data Transfer) to manage data transfer between the CPU cache and the GPU cache effi-ciently. The iCHAT technique proposed in this paper detects the communication patterns and eagerly evicts data from the owner’s caches and injects that data into the requestor’s caches. We implement the eager eviction technique of the iCHAT in a simulator based on gem5 and an AMD in-house GPU simulator. Experimental results show that the commu-nication related eviction traffic is reduced by an average of 40 % and the total directory traffic is reduced by 8 % on aver-age. 1