HPMR is a high performance computing platform based on MapReduce model. It has improved the MapReduce model to meet the need of high performance computing. Efficient memory management module ensures the efficiency of HPMR. There are two roles in HPMR, Master and worker. Master reads data blocks from the input data file and assign them to workers. Worker receives the data blocks from master, manages input and output module of the map and reduce function. The current memory management module in HPMR has some shortcomings: redundancy, inefficiency and lack of parallelism. This paper redesigned the underlying data management mechanism of HPMR based on mature memory optimization theory, proposed new memory management way based on memory pool. Experiments show that the new memory management module is necessary for efficient HPMR system.