Efficient Swap Protocol of Remote Memory Paging for Out-of-Core Multi-thread Applications

A new page swap protocol is proposed for a user-level remote memory paging system to accelerate the performance of out-of-core processing with multi-thread user programs and libraries written in OpenMP and pthread. The original swap protocol has a bottle-neck in efficient page swapping which is requested by multiple threads in a user program, because all MPI communications to memory servers and page swap managements are allocated to one system thread. Though, this protocol has a benefit that it is widely available anywhere even if MPI thread-support-level is not so high. The new protocol uses two independent system threads, one for page swapping and the other for managing user thread requests. Although it requires the highest MPI thread-support-level (multiple) which is usually considered to degrade MPI communication performance compared to than in lower MPI thread-support-level, the new protocol achieves higher performance than the original one in micro benchmark, Stream benchmark, matrix multiplication, and stencil computation.