Purdue’s Adapter for Parallel Execution and Rapid Synchronization: The TTL_PAPERS Design
Hank G. Dietz, T. M. Chung · 1995
Tightly-coupled parallel machines are being replaced with clusters of workstations connected by communication networks yielding relatively long latencies, but ever-higher bandwidth (e.g., Ethernet, FDDI, HiPPI, ATM). However, it is very difficult to make parallel programs based on fine-grain aggregate operations execute efficiently using a network that is optimized for point-to-point block transfers. TTL_PAPERS augments a cluster of PCs or workstations with the minimum hardware needed to provide very low latency barrier synchronization and aggregate communication. For example, UNIX user processes within a TTL_PAPERS cluster can perform a barrier synchronization in 2.5 microseconds, including all software overhead. This is four orders of magnitude faster than using UNIX socket connections over an Ethernet. This paper presents the TTL_PAPERS principles of operation, implementation issues, lowlevel software interface, and measured performance of the basic operations. This work has been supported in part by ONR Grant No. N0001-91-J-4013 and NSF Grant No. CDA-9015696.