While physical networks bandwidths have increased dramaticaly, these are still not usable for setting up many machines via network simultaneously. Transferring large installation images to many clients using standard unicast protocols, such as FTP or NFS, takes long hours. Therefore the most commonly used techinque of setting up a group of machines is preparing several copies of distribution media (CDs, DVDs) and performing installations independently. Most modern operating systems support non-supervised (,,kick-start'') installations to aid administrators in this task.
Another problem arises when a new application is to be installed on every node. If this application is relatively small, it can be easily installed by administrator by logging on every machine remotely and installing supplied packages. In practice such scenario is often realized using simple shell script. But if the application is big or needs some manual configuration, installation must be performed manually on every machine.
Using multicasting techniques I designed a network filesystem that is able to solve mentioned problems. After preparing one instance of completly configured filesystem (mmsnfs master), the task of populating it to all other machines (slaves) can be accomplished by simply preparing an empty disk partition, making it mmsnfs, and booting an mmsnfs-capable OS kernel. Such a mounted disk partition will behave like an NFS share, the system will boot instantly, without proper installation stage, and all transmitted files will be stored on the disk. A background daemon will download all other files, ending with individual complete on-disk mmsnfs filesystem. A reliable mechanism of updating all systems is also available.
In the rest of this paper, I discuss the goals of my work (Section 2), the principle of its operation (Section 3), current implementation state (Section 4) and some plans for my further research (Section 5).