Algorithms for supporting compiled communication

Xin Yuan, R. Methem, R. Gupta
2003 IEEE Transactions on Parallel and Distributed Systems  
In this paper, we investigate the compiler algorithms to support compiled communication in multiprocessor environments and study the benefits of compiled communication assuming that the underlying network is an all-optical Time-Division-Multiplexing (TDM) network. We present an experimental compiler, E-SUIF, that supports compiled communication for High Performance Fortran (HPF) like programs on all-optical TDM networks, describe and evaluate the compiler algorithms used in E-SUIF. We further
more » ... monstrate the effectiveness of compiled communication on alloptical TDM networks by comparing the performance of compiled communication with that of a traditional communication method using a number of application programs. while the optimizations in the communication library cannot be performed across communication patterns since the communication library does not have the information about the sequence of communication patterns in an application program. Compiled communication overcomes these limitations and exploits more optimization opportunities. In compiled communication, the compiler analyzes a program to determine its communication requirement. The compiler then uses the knowledge of the underlying network and the communication requirements to manage network resources statically. Since a typical network has limited resources and cannot efficiently support arbitrary communication patterns, the compiler must partition the program into phases such that each phase contains a fixed, pre-determined pattern that can be supported efficiently. The compiler then manages the communications within each phase statically and inserts code at phase boundaries to reconfigure the network to support the communications within each phase. Compiled communication offers many advantages over traditional communication methods. First, by managing network resources at compile time, some runtime communication overheads, such as buffer management, can be eliminated. Second, compiled communication can use long-lived connections for communication and amortize the startup overhead over a number of messages. Third, compiled communication can improve network resource utilization by using off-line resource management algorithms. Last but not the least, compiled communication can optimize arbitrary communication patterns (when the patterns can be determined at the compile time) instead of a set of predefined collective communications 2 supported by a communication library. Generally, compiled communication allows broader program optimization in comparison to library based communication since compiled communication is not limited to individual communication. The limitation of compiled communication is that it is effective only for the communication patterns that are known at compile time. Studies [17] have shown that more than 99% of communication patterns in scientific programs can be determined completely or parametrically at compile time. Only less than 1% are unknown at compile time. Thus, using the compiled communication technique is likely to improve the overall communication performance for scientific programs. All-optical interconnection networks are promising networks for future parallel computers. Multiplexing techniques, such as time-division multiplexing (TDM) [19, 25, 27] and wavelength-division multiplexing (WDM) [3], are typically used to exploit the large bandwidth in optical networks. While all-optical networks have the potential to provide large bandwidth, dynamic all-optical communication requires an all-optical path to be established before the data transmission starts. This connection management task places strict demands on the control of the interconnection network. Specifically, the network control, be it centralized or distributed, is usually performed in the electronic domain and thus is very slow in comparison to the large bandwidth supported by optical data paths. In this paper, we attempt to use the compiled communication technique to reduce the connection management overhead for the communication patterns that can be determined at compile time. We study the compiler algorithms for supporting compiled communication in multiprocessor environments and demonstrate the benefits of compiled communication on all-optical TDM networks. We present our experimental compiler, E-SUIF, which is an extension of the Stanford SUIF(Stanford University Intermediate Format) compiler [1]. E-SUIF supports compiled communication for High Performance Fortran (HPF) like programs on all-optical TDM networks. Although E-SUIF targets all-optical TDM networks, most of the compiler algorithms can also apply to other types of networks.
doi:10.1109/tpds.2003.1178875 fatcat:q3w2pf5ezbhutau5yrjd54n7fy