Performance of a Parallel Network Backup Manager

The advent of inexpensive multi-gigabyte tape drives has made possible the completely automated backup of many dozens of networked workstations to a single tape. One problem that arises with this scheme is that many computers cannot backup their disks over the network at more than a fraction of the tape's rated speed. Thus, running overnight backups sequentially can take well into the next day.

We have developed a parallel backup manager named Amanda that solves this problem by running a number of backups in parallel to a holding disk, then using a multi-buffer copy scheme to transfer the backups to the tape at the full rated tape speed. Amanda uses accurate estimates of current backup sizes as well as historical information about backup rates so as to schedule backups in parallel without swamping the network or overrunning the holding disk or tape.

Locally, we use Amanda to back up 11.5 gigabytes of data in over 230 filesystems on more than 100 workstations, using a single 2 gigabyte 8mm tape drive, taking two to three hours each night. This paper discusses the architecture and performance of Amanda.