So..the cost of an extra kernel/userspace switch was too much for a file request that is going to be serviced by a server on the other side of the internet?..really?
Putting ext3 into userspace led to slightly less than 10% performance degradation on postmark. Fuse really isn't an issue in most cases, and I can't imagine that this makes a huge difference on most workloads.
Exactly what I noticed about it. Also worth noting that the context switch here may matter for the practically tiny amount of people living in the right places in the US. It's pretty much a definition of a first world problem. Meanwhile, I'll be here with my 200ms ping to S3 wondering why would they waste all that time/complexity just to avoid a context switch.
Reminds me of the performance arguments for kdbus.
The OS pre-extension doesn't know which files are stored locally and which are placeholders for remote files. You can't just selectively use FUSE for the remote ones. Using FUSE makes every file operation slower.
Putting ext3 into userspace led to slightly less than 10% performance degradation on postmark. Fuse really isn't an issue in most cases, and I can't imagine that this makes a huge difference on most workloads.