Apply TCP-like flow control to processes accessing to the filesystem, particularly writers. That is: put them to sleep when there is "congestion", to avoid generating enormous backlogs and provide more fair allocation of disk bandwidth.

This is a nontrivial undertaking as the I/O path wasn't intended to support this type of throttling. Also, the throttle should be underneath all caching (there is nothing to be gained by throttling cached accesses) but by that point attributing I/O actions to specific processes is awkward.

It might be worthwhile to develop a scheme to tag buffers with the upper-level I/O requests their pending changes came from, or something of the sort. That would likely help with readahead processing and other things as well as with flow control.

Add a comment
Contact | Disclaimer | Copyright © 1994-2017 The NetBSD Foundation, Inc. ALL RIGHTS RESERVED.
NetBSD® is a registered trademark of The NetBSD Foundation, Inc.