ES2.1 shard recovery is extremely slow

We are upgrading to 2.1 and reindexing all the data from scratch, when growing or shrinking the cluster it takes hours to relocate or recover a shard. On the 1.7.1 cluster it usually took a few minute for a small shard, same sized shard took almost an hour to recover on 2.1. Most of the time is spent recovering translog, I tried tuning indices.recovery.translog_size it did not help.

Please see also this thread: