三值光学计算机MPI编程技术在超算集群中的使用

建立 MPI 程序中三值光学计算机(ternary optical computer, TOC)和 CPU 在指令级别上协同工作的技术. 在该技术中, TOC 节点作为服务端, 运行等待连接的进程;CPU 上运行的 MPI 进程作为客户端, 使用以SHDX 为前缀的一类扩充指令来请求与TOC 连接. TOC 和 CPU 连接之后, MPI 进程将指令级别上的服务要求和数据一并发送给TOC 节点, TOC 发挥其处理器可构成复合运算器及数据位数众多的优势来处理数据, 并将计算结果回送给客户端的MPI 进程. 该过程实现了 CPU 在指令级别上调用TOC 来协同工作的目的. 目前, 设定一个超级计算机系统可以融入10 个TOC 节点, 每个TOC 节点可以打开8 个服务端口, 未来的版本可能扩大这两个限制. 实验结果证明了该技术的正确性.

MPI Programming Based on Ternary Optical Computer in Supercomputer

An MPI method is proposed to make the ternary optical computer (TOC) node cooperate with other general compute nodes in a cluster supercomputer in an MPI program. The TOC nodes work as servers, running a process to wait for connections. Any MPI process running on CPU works as a client side, using a series of expand instructions with the prefix of SHDX to contact the TOC node. After the TOC node gets contact with the MPI process on the CPU nodes, MPI process sends an operation request and data to the TOC in a connection. Then, TOC exerts the advantages of its reconfigurable processor and high data-bit-width to process the data and sends the results back to the MPI process. This method realizes using TOC on the instruction level. At present, there can be at most 10 TOC nodes in a supercomputer system, and each TOC node opens 8 ports. Future versions may break these restrictions. Experimental results show feasibility and practicality of the method.