DCC: A Replacement Strategy for DBT System Based on Working Sets

Limited memory resource has always been deemed as the major bottleneck of the program performance, not excepting for dynamic binary translation (DBT) [1] system. However, traditional methods seldom enable this issue mentioned to be solved better due to that a fix cache size routinely assigned for codes without considering the program behavior on the fly. Though some prediction methods implemented via LRU [2] histograms can achieve better performance in Operating System, unlike pages in memory, blocks in DBT have unfixed sizes each other, making the prediction imprecise. In this paper, we present a new replace policy named DCC (Dynamic Code Cache) based on working set [3] for bounded code cache, which would dynamically change the size of the code cache and clear the cache according to the information of working set detected. It would sacrifice a small part of the performance to save the space, especially when many great or complicated applications are running on the same physical machine, and the performance might be better under this condition. For evaluating, we test this strategy on a DBT system named Crossbit [4]. We get around 30% space saved up with about 12% running time sacrificed.