Abstract: There are two bottle-necks in scalable parallel processing systems:（1）Thetransmission delay in interconnection network, and （2） The throughput capability.According to the characteristics of SPP architecture and practical requirements, this paperproposes to add special I/O processors between servers and SM/SSM to realize highspeed transfering, thus improving I/O performance.This method not only eliminatesthe transmission delay, but also supports parallel execution. between processors andDMA controllers.