I update large collection (~500gb snappy compressed data) on non-sharded environment. I split collection into chunks using splitVector, 32mb per chunk. Then I perform some analysis on each record and update each record (add one small field). At the start everething is cool, performance comparable to TokuMX which I used earlier. But then performance degrades rapidly. You can find plots of some metrices attached. It is very strange that disk is not fully utilized, but cpu utilized at 100%.
My db hosted on ec2 r3.4xlarge machine. Disks - 2 * 1TB ebs combined into raid0.