I see that Dana H. Myers has been digging away at ZFS performance using the only metric that really matters to the real geek: OS build performance. The numbers are interesting... I’m surprised that compression made so little difference, both source and object code compress quite well (I just ran a little test: the Emacs binary compressed to 18% of its size, a bunch of Java code to 19%.) Maybe the fact that it’s zillions of little files means that the file open/create overhead dominates the actual input/output time? There is no doubt there is a huge amount of work to be done on I/O performance, both understanding it and improving it. But ZFS is increasingly looking like a step forward.
