Measurements regarding data backup performance with the backup file located on the Logical Volumes of a ETERNUS CS8200 VTL were performed on a Server SE710. The results also apply to an SU730. The storage system volumes that were used with HSMS BACKUP- and RESTORE-FILES contained files varying in size (regarding PP, PAM Pages):
- Small files: 35510 * 621 PP (1,2 MB)
- Medium-sized files: 3630 * 6075 PP (12 MB)
- Large files: 36 * 621600 PP (1,2 GB)
- Systemmix: 11872 files of different sizes
The following hardware configuration was used:
- Server: SE710-20D (1220 RPF)
- Storage-System: ETERNUS DX8700, 16 Gbit/s FC, 2,5" SSD-M Drives
Volumes: NK2 format, in 2 RAID-Groups (RAID1), 4-way connection - ETERNUS CS8200 VTL: FC-connection with 16 Gbit/s via 2 channels, Logical Drives IBM 03590E1A (BS2000: DEVICE-TYPE=TAPE-C4)
Backup throughput with ETERNUS CS
The backups were performed on one Pubset with 1 task on 1 Logical Volume as well as using parallel backups of 2 Pubsets with 2 tasks on 2 Logical Volumes. With FastDPAV the following average throughputs were measured:
| HSMS/ARCHIVE throughput (MB/s) | |||
BACKUP | RESTORE | BACKUP | RESTORE | |
small files | 317 | 198 | 620 | 282 + 105 |
medium files | 317 | 253 | ||
large files | 317 | 254 | 633 | 473 |
The throughputs using RESTORE are lower than those achieved using BACKUP. The Cache Hit Rate Write of the storage system occasionally dropped far below 100% during RESTORE operations. RESTORE subtask 1 with 2 small files and 2 tasks used about double the time of subtask 0. Hence in the table above values for 2 parallel tasks as well as the run-time of subtask 1 are pointed out. The many access operations on the BS20000-catalog lead to a significant overhead.
Comparison to ETERNUS DX
Comparison measurements with the HSMS backup file located on a Pubset of the storage system were conducted. The Pubset was connected via a different 4-way connection than the backed up Pubsets.
The throughputs achieved with FastDPAV as well as without PAV were as follows:
| HSMS/ARCHIVE throughput (MB/s) | |||
BACKUP FastDPAV | RESTORE FastDPAV | BACKUP without PAV | RESTORE without PAV | |
small files | 290 | 158 | 170 | 111 |
medium files | 471 | 515 | 176 | 187 |
large files | 463 | 477 | 177 | 197 |
In this scenario with FastDPAV the BACKUP throughput using medium and large files is about 50% higher compared to using Logical Volumes. For RESTORE operations the throughput is about twice as high as before. The parallel Read- and Write-IOs on the backup file impact the results heavily.
Scaling with several tasks
When connecting the Logical Drives via 1 channel to the server the BACKUP operations with several parallel tasks scale well. When using RESTORE the storage system ist the limiting factor regarding throughput. HSMS-Backups (FastDPAV, volumes with large files) with a single task and 2/3/4 parallel tasks resulted in the following BACKUP and RESTORE throughputs:
| HSMS/ARCHIVE throughput (MB/s) | |||
BACKUP TAPE-C4 | RESTORE TAPE-C4 | BACKUP TAPE-U4 | RESTORE TAPE-U4 | |
1 task | 317 | 254 | 316 | 253 |
2 tasks | 625 | 467 | 621 | 465 |
3 tasks | 917 | 493 | 911 | 491 |
4 tasks | 1089 | 464 | 1104 | 464 |
The Logical Drives were configured as both IBM 03590E1A (BS2000: DEVICE-TYPE=TAPE-C4) and LTO Ultrium 4 (BS2000: DEVICE-TYPE=TAPE-U4).
There was no significant difference between DEVICE-TYPE TAPE-C4 and TAPE-U4.
Comparison K- and NK-Pubsets
Conducting further measurements the difference between backing up K-Pubsets and NK-Pubset with the backup file located on Logical Volumes was examined. The Pubsets each contained files of different sizes (Systemmix). Both BACKUP and RESTORE were used and tested several times. The following table describes the average throughputs for the different scenarios:
| HSMS/ARCHIVE throughput (MB/s) | |||
BACKUP FastDPAV | RESTORE FastDPAV | BACKUP without PAV | RESTORE without PAV | |
K - K | 287 | 175 | 124 | 93 |
K - NK | 287 | 11 | 124 | 11 |
NK - NK | 291 | 176 | 254 | 162 |
NK - K | 290 | 178 | 249 | 109 |
The results varied siginificantly between runs, most of all while using RESTORE. With BACKUP K and a following RESTORE NK all files were converted using PAMINT. This lead to a significant throughput decrease to 11 MB/s. This decrease in performance is unrelated to PAV, but is the reason for increased RESTORE run-times.