elasticsearch cluster index performance VS single node index performance

[reason001]

Dear all: I found that single data node indexing speed is much more faster than multiple data node cluster. Now I am setuping Elasticsearch cluster with 3 nodes, 1 index and 9 shards. One node is empty data node, acting as LB server, and other two are data node. When I using Jmeter to test index performance, I found that the thoughput is very low, less than 100/sec. But if I shutdown one of the data node, test with one LB server and one data node, the throughput is rising up to more than 1000/sec. I've used some advice found in forum, like increase heap size and increate refresh interval, but not helpful, the two data node solution is always having bad performance compares to two data node solution. Not sure if any one saw such situation before, please check if you can give me some advice, thanks.