异构千核处理器系统的统一内存地址空间访问方法

为了达到异构多核处理器能直接交叉访问对方的内存地址空间的目的,通过构建统一的三级Cache结构和数据块状态标记方法,并优化Cache块状态的修改算法,提出了异构千核处理器系统的统一内存地址空间访问方法,避免了当前独立式异构计算机系统结构下复制和传输数据块所带来的大量额外访存开销。通过采用部分Rodinia基准测试程序测试,获得了最高9.8倍的系统加速比,最多减少了90%的访存频率。因此,采用该方法能有效减少异构核心间交换数据块所带来的系统开销,提高异构千核处理器的系统性能加速比。

An approach to accessing unified memory address space of heterogeneous kilo-cores system

In order to access independent memory space of CPU and GPU directly from opposite directions, an effective approach to accessing unified memory address space of heterogeneous kilo-cores system is proposed, which is implemented by building a unified level-3 Cache and tagging blocks in Cache, and optimizing the algorithms of modifying the states of blocks. Therefore, the heterogeneous kilo-cores system avoids significant overhead of accessing memory instead of that in current discrete hybrid computer system equipped with GPUs by PCI-E. According to the results of experiments from partial programs of Rodinia benchmarks, a maximal speedup by 9.8x and maximal decrease of load/store instructions by 90% are gained. In conclusion, it's certified that our solution is effective to decrease overhead of transferring data among computing units in heterogeneous system and significantly enhance the whole system computing performance.