AMANDA’s limitation is that it cannot span tapes. As we are now backing up 60 GB partitions that ar 90% full, this leads to an inefficient use of tape space - the extra space at the end of a 40/80 GB tape can’t be used as there are no small filesystems to put there. As a result, it takes more than 8 tapes to make a full backup of the entire server set where it should take no more than 5.
NetBackup is resident on a new server: EYEWI, a SunFire V240 running Solaris 9. We have moved the tape library to this server and have also attached the DLT40000 drive that we had gotten with PAX originally. The main backups are to the library while the NetBackup catalog backups are to the DLT40000.
NetBackup requires a distinctly different mindset than AMANDA in its conception of backup schedules and retention times. Where AMANDA uses a tape cycle concept to implement a retention period (that is, you have _N_ tapes with a cycle of _M_ days, thereby deriving a full backup frequency and retention period), NetBackup is perhaps less intelligent. Instead, you define the frequency and windows for full and incremental backups, and you also define an explicit retention period for these backups. This, together with a client and file list makes up the NetBackup concept of a policy.
We have implemented NetBackup policies such that there are at most three policies per client and no two clients share a policy. While it is potentially possible to put multiple similar clients into a single policy, we have few enough clients that are this similar and there is a management simplicity of one client in a policy. We have defined policies for user, system, and business data. Some clients may not have some of these (for example, the WebDB server has no user data, and several of the Solaris servers have system data that is easily replicable from a Jumpstart). The user and business data are written to tape, and the system data is compressed and saved on a disk storage unit on EYEWI.
The catalog database is backed up to local disk after every successful policy backup. The post-backup script has been modified to make a snapshot of the filesystem and dump it to the DLT40000 tape.
Solaris and NetBackup require some tuning for optimum performance. Veritas has this document on Solaris kernel tuning for shared memory and IPC tuning and this document on NetBackup shared memory and network performance tuning - the result is better throughput on the network and tape systems. Putting EYEWI on gigabit ethernet also speeds up backups.
Yesterday I rebuilt the main campus PacketShaper to solve some lingering problems we've been seeing with a number of protocols.
I think this came from upgrading the firmware earlier this fall. We were seeing problems with FTP transfers in particular, although I had noticed problems with SSL versions of IMAP and SMTP, and others had been having (possibly unrelated) problems with the library proxy system. Resetting the shaper made it crash twice on every boot with an invalid configuration error. Since the error was unhelpful in the extreme, I opted to completely rebuild the shaper from factory defaults.
Everything seems to be working properly, again. A side benefit of this is that we can clean out the cruft that has accumulated over the past couple of years.