Summary

As expected, the simplicity of physical backup and restore means that they can achieve much higher throughput than logical backup and restore which are constantly interpreting and creating file system meta data. The performance of physical dump/restore scales very well; when physical restore hits a bottleneck it is simple to add addition tape drives to alleviate the bottleneck. Eventually the disk bandwidth will become a bottleneck, but since data is being read and written essentially sequentially, physical dump/restore allows the disks to achieve their optimal throughput. Logical dump/restore scales much more poorly. Looking at the performance of 4 parallel logical dumps to 4 tape drives (Figure 5) we see that during the writing files stage the CPU utilization is only 90% and the tape utilization is under 70% (as compared to that achieved by physical dump). The bottleneck in this case must be the disks. The essentially random order of the reads necessary to access files in their entirety achieves highly sub-optimal disk performance.