Solved Paper: To FUSE or Not to FUSE: Performance of User-Space File Systems


Son of Beastie

Reaction score: 2,351
Messages: 3,241

Note that Bharath and Vasily measured the performance of FUSE on Linux. Their results might apply to FUSE on other operating systems, but we should expect significant performance difference, because some fine details of how FUSE interacts with the kernel matter greatly. I general their result can be described as: Using FUSE has an overhead, but it is moderate and can be near zero for some workloads; perhaps most relevant is that FUSE eats extra CPU.

But: The results explicitly do not generalize to other file systems that are implemented in userspace without using FUSE. There are lots of examples, some of which they list in the paper (GPFS, DDFS, GFS, Nimble, HDFS), and some they don't describe (for example Los Alamos supercomputer file system adaptor). Some of those file systems are among the fastest in the world. The generalization "userspace file systems are slow" is definitely wrong, and not implied by their paper.