A memory schedule policy oriented to stream architecture
Chiyuan Ma, Xiaoqiang Ni · 2014
As the gap between processor speed and memory speed continues to increase, memory system becomes the performance bottleneck of the processor. The memory system of stream architecture is different to that of general architecture, how to speed up the stream data access is a challenging issue for current stream architectures. This paper proposes a new memory schedule policy oriented to stream architecture. First, a new data buffer structure is used to buffer stream data so that multiple requests of stream data accesses can be merged in request queue and data buffer, then by collecting statistical information of address distribution of all the requests waiting for accessing memory and analyzing the regular changes of access addresses, this policy can predict the access behavior of each bank and uses appropriate page policy to each DRAM bank. The experimental results show that our policy can improve the performance by 5.5% comparing with the commonly used page policies.