Abstract: The booming of many-core architecture and sharing data model applies new challenge to memory architecture. Neither traditional shared last level cache nor traditional private last level cache performs well in new situation. Cooperative Caching performs better in many-core architecture via incooperating advantages of both architectures. However, when facing many-core architecture, long-distance inter-core accesses, resulting from its unique reservation of victim line, greatly increases on-chip communication and leads to performance loss. Thus, this paper proposes an adaptive data sharing mechanism for private last level cache which allows multiple reservations of victim lines, with the degree of reservation adaptively adjusted. Simulation results show that compared with Cooperative Caching, the proposed adaptive data sharing mechanism reduces on-chip communication by 12.8% at average and by 32.7% at best; overall performance is improved by 9.1%. The result proves the proposed architecture performs better under the situation of many-core architecture and sharing data model.

Key words: many-core architecture, last level cache management, victim line reservation