面向大数据应用的众核处理器缓存结构设计

大规模数据排序、搜索引擎、流媒体等大数据应用在面向延迟的多核/众核处理器上运行时资源利用率低下,一级缓存命中率高,二级/三级缓存命中率低,LLC容量的增加对IPC的提升并不明显。针对缓存资源利用率低的问题,分析了大数据应用的访存行为特点,提出了针对大数据应用的两种众核处理器缓存结构设计方案,两种结构均只有一级缓存,Share结构为完全共享缓存,Partition结构为部分共享缓存。评估结果表明,两种方案在访存延迟增加不多的前提下能大幅节省芯片面积,其中缓存容量较低时,Partition结构优于Share结构,缓存容量较高时,Share结构要逐渐优于Partition结构。由于众核处理器中分配到每个处理器核的容量有限,因此Partition结构有一定的优势。

Cache structure design for big data oriented many-core processor

Some big data applications such as data sorting, search engine, streaming media running on the traditional latency oriented multi/many core processor are inefficiency. The hit rate of L1 Cache is high while that of L2/L3 Cache is relative low and IPC is not sensitive to LLC capacity. To address the low utilization issue of cache resources, we analyze the memory access patterns of big data applications, and then propose an optimization method of cache structure for many core processor. Both the two structures  only have L1 cache, while one is fully shared cache structure, and the other is partly shared cache partition structure. The evaluation results show that these two schemes can significantly save chip area at the cost of slightly increase of memory access. When cache capacity is low, the partition structure is superior to the share structure. As cache capacity increases, the share structure will gradually become superior to the partition structure. For many core processors, the capacity assigned to each processor is limited, thus the partition structure has certain advantages.