DownloadGlusterFS cookbook
From GlusterDocumentation
FIXME: work in progress - suggestions welcome
Contents |
Introduction
What is GlusterFS
A largely scalable clustered filesystem in userspace (FUSE). GlusterFS is highly adaptable, feature rich, POSIX compliant and has a layered design. GlusterFS can work on any type of interconnect, IB or gig/e, or 10gig/e.
Why we wrote another filesystem
We ourself wanted to deploy a large filesystem for one of our customer, who needed Petabyte + storage. For that type of large volume, we needed a filesystem which addresses reliability, maintainability (ease of use), and scalability. We couldn't find all these three in a single filesystem. This is the reason why we started our own filesystem.
Where is white paper about your filesystem format
GlusterFS is a clustered network filesystem, not a disk based filesystem. with GlusterFS, on back end one can keep their own pet filesystem in the back end directory. GlusterFS doesn't have any format of its own to keep data. A back end of GlusterFS can be visualized same as NFS export point, where you can keep the directories as it is. As its not a new filesystem format by itself, we happen to spend more time coding than publishing a white paper. But we will come out with one soon until then the answer will be "No, we don't have one yet!".
Why use GlusterFS
As told earlier, we wrote it to address three major issues seen in large scale storage solutions.
- Scalability
- Performance
- Manageability
The Userspace design approach which we took helped us to achieve these goals faster, in a much efficient way.
Its well known that filesystems are core part of the OS, hence has to be in kernel space. But when it comes to network filesystems, this is is not true. The delay which is caused by network latency is much larger than context switch overhead caused by being in userspace. And if the network latency is reduced by using Infiniband or 10gigabyte cards, one can do RDMA from userspace to remote machines, hence the question of context switch doesn't arise. So, the question of performance doesn't come into picture.
Also as we are in userspace, the development cycles for any feature is quite less compared to kernel based filesystem. Any problem with the filesystem, just a application restart is enough (say a umount/mount is enough). No need of reboots, no nightmares of kernel panics.
