CSDA: An Adaptive Data Storage Strategy for Predictive Branch Path on Multi-Core Architecture

Abstract

“Memory Wall” has already become the chief barrier for processor performance promotion. Cache pollution caused by speculative execution of memory instructions in predictive path may impact the performance of processor seriously. This paper proposes an adaptive storage frame to filter pollution data from predictive path based on confidence estimation mechanism for multi-core architecture, called CSDA. Firstly, CSDA makes use of confidence estimation mechanism to separate out the more likely mis-predicted paths from all predictive paths adaptively. Secondly, CSDA divides all memory reference instructions in low confidence predictive path into pre-fetching type and pollution type according to the low confidence pollution type memory reference recognition history table. Thirdly, a special low priority Load/Store queue is created for these pollution type memory reference instructions, and pollution data cache is used to store pollution data. Finally, CSDA may alleviate the negative effect of cache pollution data, and promote the effective memory bandwidth. Simulation result indicates that, in dual-core configuration, CSDA could avoid most cache pollution and upgrade performance. Relative to the baseline architecture, miss rate of D-Cache reduces from 9% to 23%, average 17%. Miss rate reduction of L2 cache ranges from 1.02% to 14.39%, average 5.67%. IPC improves from 0.19% to 5.59%, average 2.21%.