SkyeFS is a FUSE filesystem providing distributed directories using Giga+ and PVFS.

There are a huge variety of distributed filesystems available and a great deal of research has gone into the problem of distributing filesystem data over a cluster of machines. However, many of these filesystems are unable to scale metadata. Some systems, such as HDFS, store all directory entries on a single machine while others, like PVFS, store all entries in a particular directory on a single machine. Both of these techniques scale poorly for directories with many entries or highly concurrent write workloads.

Giga+ is a technique for distributing the directory entries in a directory between a set of servers. When a directory is initially created, it is entirely contained in one server. As the directory grows, it splits and the hash space of file names is distributed to more servers until the load is balanced between the available servers.

SkyeFS uses the Parallel Virtual File System (PVFS) as the underlying distributed filesystem on top of which we layer the Giga+ scheme. We use multiple PVFS directories to represent each logical SkyeFS directory and a client and server pair to manage the partitioning. The Filesystem in Userspace tool provides us with an easy way to implement a VFS compatible filesystem without writing kernel code.

Using SkyeFS on top of PVFS allows your directories to scale with the number of available servers. If you need to store millions of files in a single directory or are doing highly concurrent inserts and deletes in a large directory, you may see improved performance and server utilization.

The SkyeFS client and server are lightweight wrappers around PVFS operations and, other than the FUSE overhead, add very little additional overhead to PVFS operations. SkyeFS was designed to ensure that server failures will never leave the filesystem in an inconsistent state, maintaining the failure semantics of PVFS.

Yes! But, depending on your use case, some more work may be required before you'll be happy with it. Currently, there are same areas with unfinished work. This includes some of the failure recovery code (e.g. resuming partition splits) and some performance sensitive code (e.g. flow control in the server). These areas are indicated with TODO or FIXME comments in the code.

First, setup an empty PVFS filesystem. Then, launch an instance of the skye_server on each PVFS MDS. You can now use skye_client to mount the filesystem. Run these commands without arguments for their usage. Or see the scripts in util/ for examples.