LWN: Comments on "The ZUFS zero-copy filesystem"

The ZUFS zero-copy filesystem

mtanski — Thu, 07 Jun 2018 15:09:22 +0000

Can this be leveraged to enable faster netfs in userspace?

As a remote filesystem you pretty much need to be in the kernel proper to get any kind of resonable performance. FUSE works if you're in a pinch, but not performance needed for any kind of production work. Some of the netfs can perform their transfers over RDMA.

I know this is not the case environed, some filesystems might still need a caching on the client side. But there's no reason a caching solution cannot be developed in userspace and then shared via memory in a zero-copy like fashion.

The current requirement of putting the netfs in the kernel has a lot of costs in terms of maintance and just general complexity.

The ZUFS zero-copy filesystem

willy — Wed, 06 Jun 2018 22:41:45 +0000

Apart from the handwaving bits that don't actually work ;-)

The ZUFS zero-copy filesystem

jlayton — Wed, 06 Jun 2018 20:37:16 +0000

IMO, the interesting bit here is the zero-copy, fast transport between userland and kernel. Layering this on top of pmem is neat (and fits Boaz requirements), but a lot of the concepts that Boaz is using here could make for a potential foundation for a better-performing replacement for FUSE.

The ZUFS zero-copy filesystem

abacus — Wed, 06 Jun 2018 18:14:38 +0000

Why would anyone use ZUFS since there is a persistent memory filesystem available that is simpler and that performs significantly better, namely NOVA?