Elasticsearch threads behaves different from each other
Some Threads do more work than others. The one(s) doing network communication always generate load, threads doing querys may generate high load and so on. As far as i know, lucene merges are quite CPU hungry.