Basic Backup / Restore

Table 2 reports our measurements of backing up and restoring a large, mature¹ data set to a single DLT-7000 tape drive. In this configuration both logical and physical dump obtain similar performance, with physical dump getting about 20% higher throughput. This is because the tape device that we are using (a DLT-7000) is the bottleneck in the backup process. Note however the significant difference in the restore performance. This can be primarily attributed to image restore's ability to bypass the file system and write directly to the disk, while logical restore goes through the file system and NVRAM².

   

Table 2: Basic Backup and Restore Performance

Operation	Elapsed time (hours)	MBytes/second	GBytes/hour
Logical Backup	7.4	7.2	25.0
Logical Restore	8.1	6.3	22.8
Physical Backup	6.2	8.6	30.3
Physical Restore	5.9	8.9	32.0

   

Table 3: Dump and Restore Details

Stage	Time spent	CPU Utilization
Logical Dump
Creating snapshot	30 seconds	50%
Mapping files and directories	20 minutes	30%
Dumping directories	20 minutes	20%
Dumping files	6.75 hours	25%
Deleting snapshot	35 seconds	50%
Logical Restore
Creating files	2 hours	30%
Filling in data	6 hours	40%
Physical Dump
Creating snapshot	30 seconds	50%
Dumping blocks	6.2 hours	5%
Deleting snapshot	35 seconds	50%
Physical Restore
Restoring blocks	5.9 hours	11%

We can look at the resource utilization of the filer while backup and restore are proceeding to learn something about how logical and physical backup are likely to scale as we remove the bottleneck device. Table 3 indicate the time spent in various stages of backup and restore as well as the CPU utilization during each stage. It is worth noting the variation in CPU utilization between the two techniques. Logical dump consumes 5 times the CPU resources of its physical counterpart. Logical restore consumes more than 3 times the CPU that physical restore does.

One way to reduce the time taken by the backup procedure is to backup multiple file system volumes concurrently to separate tape drives. The resource requirements of both logical dump and physical dump are low enough that concurrent backups of the home and rlse volumes did not interfere with each other at all; each executed in exactly the same amount of time as they had when executing in isolation.