A high performance barrier synchronizer and its novel applications in highly parallel systems
Kanad Ghose, Der-Chung Cheng · 2002
Conventional approaches for implementing barrier synchronization in large scale systems result in long synchronization times due to memory and network contention. The authors present some applications of a very simple barrier hardware that allows barrier synchronizations to be performed within a few processor cycles. The low latency of this barrier synchronizer and its ability to allow dynamic participation permits several rather unconventional applications like low-latency software combining, scans and determination of extremum values in constant time. This paper presents these applications.>