SCI Networking for Shared-Memory Computing in UPC: Blueprints of the GASNet SCI Conduit H. Su, B. Gordon, S. Oral, A. George High-performance Computing and Simulation (HCS) Research Lab, Dept. of Electrical and Computer Engineering, University of Florida, Gainesville, Florida 32611-6200 {su, gordon, oral, george}@hcs.ufl.edu Abstract Unified Parallel C (UPC) is a programming model for shared-memory parallel computing on shared- and distributed-memory systems. The Berkeley UPC software, which operates on top of their Global Addressing Space Networking (GASNet) communica- tion system, is a portable, high-performance implementation of UPC for large-scale clusters. The Scalable Coherent Interface (SCI), a torus-based system-area network (SAN), is known for its ability to provide very low latency transfers as well as its direct support for both shared-memory and message-passing communications. High-speed clusters constructed around SCI promise to be a potent platform for large-scale UPC applications. This paper introduces the design of the Core API for the new SCI conduit for GASNet and UPC, which is based on Active Messages (AM). Latency and bandwidth data were collected and are compared with raw SCI results and with other existing GASNet conduits. The outcome shows that the new GASNet SCI conduit is able to provide promising performance in support of UPC applications. Keywords - Scalable Coherent Interface, Global Ad- dress Space Networking, Unified Parallel C, Active Messages. 1. Introduction Many scientific as well as commercial endeavors rely on the ability to solve complex problems in a quick and efficient manner. One of the dominant so- lutions to this problem has been the advent of parallel computing. To supplement the architectural im- provements in this area, parallel programming models have emerged to provide programmers alternate ways in solving complex and computationally intensive problems. Such models include message passing, shared memory, and global address space. While message passing and shared memory are the two most popular ways to implement parallel programs, global address space is quickly gaining momentum. One of the reasons for this development is the growing acceptance of Unified Parallel C (UPC) [1,2] and other models like it. UPC is a parallel extension to the ISO C standard that gives programmers the ability to create parallel programs that can target a variety of parallel architecture platforms while maintaining a familiar C-style structure. This approach allows a smaller learning curve for people with C experience to begin creating parallel programs and often results in tighter and more efficient code. One recent development in UPC is the interest in providing a means for executing UPC over commer- cial-off-the-shelf (COTS) clusters. The Berkeley UPC runtime system [3], developed by U.C. Berkeley and Lawrence Berkeley National Laboratory (LBNL), is a promising tool now available to support this en- deavor. An underlying key to this system is the Global Addressing Space Networking (GASNet) communication system [4,5]. GASNet defines a standard application interface that can be implemented over a wide variety of standard and high-performance networks such as Ethernet, InfiniBand, Myrinet, and Quadrics. In this study, we present the design of a new GASNet conduit operating over the Scalable Coherent Interface (SCI) network [6]. Benchmarks were exe- cuted on the newly developed conduit and compared against the raw performance of SCI, the GASNet Myrinet conduit, and GASNet MPI conduit on SCI using Scali’s ScaMPI [7] to evaluate various strengths and weaknesses. The next section of this paper briefly describes the architecture of SCI and GASNet. In Section 3, we discuss related research. Section 4 describes the design