MapReduce　is　an　effective　programming　model　for　analyzing　large-scale　data.　Hadoop-a　distributed　processing　system　is　widely　used　nowadays.　Improving　the　task　parallelism　can　be　a　key　point　to　improve　the　MapReduce　performance　in　Hadoop.　In　this　paper,　we　address　the　problem　in　two　ways.　On　the　one　hand　we　can　run　the　tasks　with　some　dynamic　configurations.　On　the　other　hand,　considering　of　the　difference　of　tasktracker　we　use　mathematics　method　to　predict　the　cups＇　utilization　of　tasktracker　to　assign　the　task.　Experimental　results　on　both　ways　show　we　can　improve　the　performance　in　Hadoop　by　improving　the　task　parallelism.