The parallelCollectionScan command pre-fetches the first extent for each cursor upon construction of the cursors. During cursor iteration, when a cursor needs to hop to the next extent, it also pre-fetches that extent. This can result in pre-fetching of large amounts of data. Performance tests show that disabling pre-fetching is more performant overall in most cases.
For large numbers of cursors with a big collection, this can mean aggressive pre-fetching of many gigabytes. This can cause large, unwanted spikes in disk IO.
All production releases up to version 2.6.3 are affected by this issue.
The fix is included in the 2.6.4 production release.
Pre-fetching was disabled for the parallelCollectionScan command.