Distributed shared memory architecture for JUMP-1 a general-purpose MPP prototype

We describe and evaluate a novel distributed-shared memory (DSM) architecture of JUMP-1, a general-purpose MPP system. For improving performance, JUMP-1 DSM architecture utilizes cooperation of the network, construction of memory-directories and a memory-protocol that unifies memory-consistency, communication and synchronization. Among features of JUMP-1 DSM, we show details of Reduced Hierarchical Bit-map Directory schemes (RHBDs) which utilize hierarchy embedded in the interconnection network for reducing network traffic on shared memory operations. Three variations of the RHBD are implemented on a network called the RDT (Recursive Diagonal Torus) consisting of the hierarchical structure of two-dimensional tori. In RHBDs, the bit map directory is reduced for quick multicasting without accessing the directory in each hierarchy. Most unnecessary packets caused by reduction of the bit map are removed with the pruning cache provided in the router. The results of simulation demonstrate that latency for cache coherent messages is much improved compared with traditional directory schemes.

[1]  A. Agarwal,et al.  An evaluation of directory schemes for cache coherence , 1988, [1988] The 15th Annual International Symposium on Computer Architecture. Conference Proceedings.

[2]  Kei Hiraki,et al.  Complementary hybrid architecture with two different processing elements with different grain size , 1994, Proceedings of TENCON'94 - 1994 IEEE Region 10's 9th Annual International Conference on: 'Frontiers of Computer Technology'.

[3]  Seif Haridi,et al.  Data Diffusion Machine - A Scalable Shared Virtual Memory Multiprocessor , 1988, FGCS.

[4]  Kai Li,et al.  IVY: A Shared Virtual Memory System for Parallel Computing , 1988, ICPP.

[5]  Erik Hagersten,et al.  DDM - A Cache-Only Memory Architecture , 1992, Computer.

[6]  Anant Agarwal,et al.  Directory-based cache coherence in large-scale multiprocessors , 1990, Computer.

[7]  Anoop Gupta,et al.  SPLASH: Stanford parallel applications for shared-memory , 1992, CARN.

[8]  Tomoyuki Tanaka,et al.  MISC: A Mechanism for Integrated Synchronization and Communication Using Snoop Caches , 1991, International Conference on Parallel Processing.

[9]  Hideharu Amano,et al.  The JUMP-1 router chip: a versatile router for supporting a distributed shared memory , 1996, Conference Proceedings of the 1996 IEEE Fifteenth Annual International Phoenix Conference on Computers and Communications.