CDP: A multithreaded implementation of a network communication protocol on the Cyclops-64 multithreaded architecture

A trend of emerging large-scale multi-core chip design is to employ multithreaded architectures---such as the IBM Cyclops-64 (C64) chip that integrates large number of hardware thread units, main memory banks and communication hardwares on a single chip. A cellular supercomputer is being developed based on a 3D connection of the C64 chips. This paper introduces our design, implementation, and evaluation of the Cyclops Datagram Protocol (CDP) for the IBM C64 multi-threaded architecture and the C64 supercomputer system. CDP is inspired by the TCP/IP protocol and its design is very simple and compact. The implementation of CDP leverages the abundant hardware thread-level parallelism provided by the C64 multithreaded architecture.; The main contributions of this paper are: (1) We have completed a design and implementation of CDP that is used as the fundamental communication infrastructure for the C64 supercomputer system. It connects the C64 back-end to the front-end and forms a global uniform namespace for all nodes in the heterogeneous C64 system; (2) On a multithreaded architecture like C64, the CDP design and implementation effectively exploit the massive thread-level parallelism provided on the C64 hardware, achieving good performance scalability; (3) CDP is quite efficient. Its Pthread version can achieve around 90% channel capacity on the Gigabit Ethernet, even it is running at the user-level on a single processor machine; (4) Extensive application test cases are passed and no reliability problems have been reported.