FUSE root fs, embedded, performance

Posted Mar 20, 2009 4:32 UTC (Fri) by szaka (guest, #12740)
Parent article: Unioning file systems: Architecture, features, and design choices

FUSE can be used in many different ways. The below comments are just some short and minor corrections. It's a totally different issue how much they could be true for a FUSE-based union file system.

Thankfully FUSE file systems can be used as root file systems. For instance NTFS-3G is used by some Linux distributions as root file system for about two years.

FUSE is often used in embedded systems. One example is djmount but NTFS-3G is also included in many consumer electronics, NAS, set-top boxes, multimedia players/recorders and many other different type of devices.

Typically the dominant performance factors of general purpose, block device based file systems are the file system design, quality of the implementation and optimization/tuning to the specific hardware platform. A FUSE file system can be high-performance too if it's used efficiently.

FUSE can be considered as a kernel file system driver where the slow paths are forwarded to user space. One could also think about it as a high-performance kernel network file system with extremely low latency and high bandwidth. For instance the current best write performance of the NTFS-3G driver is 1.83 GB/s and it's the 3rd fastest in metadata operations after btrfs and ext4. Which I think is not too bad considering that it's not fully optimized yet.

FUSE root fs, embedded, performance

Posted Mar 20, 2009 5:52 UTC (Fri) by vaurora (guest, #38407) [Link]

Thanks for the corrections! I usually learn as much from the user comments as from researching the article and this story is no exception.

Perhaps I'm just paranoid (well, I know I am), but I still don't want access to my root file system to go through any sort of userspace intermediary. I am glad to know that it is possible.

"Embedded" covers a lot of ground. I worked on an "embedded" system with 1GB of RAM back in 2002. Whenever I think that even the tiniest devices have megabytes of RAM, someone proves me wrong again...

I don't discuss performance in any detail any more. :) No matter what you say, someone else's workload performs better or worse, and the software can always be more optimized.