Chart 9 - Database Backup To 15/30 Dlt Tape Raid Using Arcserve - Compaq 219700-001 - ProLiant - 1500 White Paper

Compaq Backup and Recovery for Microsoft SQL Server 6.x

Chart 9 - Database Backup to 15/30 DLT Tape RAID Using ARCserve

As we can see, using the ARCserve software in conjunction with SQL Server yields excellent
throughput and scalability, both with RAID-0 and RAID-5 tape arrays. The average increase in
throughput when adding a tape drive to either type of array is around 4.25 GB/hr. As with disk arrays,
RAID-5 tape performance falls below RAID-0 performance; the average difference between the two is
about 5.3 GB/hr. In fact, the throughput for an array of 'n' tape drives in a RAID-5 array is close to the
performance of 'n-1' drives in a RAID-0 array. This makes sense when we consider that for every 'n'
blocks of data (1 per drive) sent to the RAID-0 array in a single "stripe", 'n-1' blocks of data and 1
block of parity are sent to the RAID-5 array. There is some additional performance overhead in RAID-
5, probably because the program has to calculate the parity block (a boolean XOR operation on the
data blocks) and interleave it among the drives. Although not shown in the chart, the RAID-5
operations had a CPU% usage that was slightly more than for the RAID-0 operations, and a PCI%
usage that was slightly less. The fault tolerance provided for the data set with the RAID-5 option
however, may be worth the performance penalty if archiving data in a mission-critical environment.
The next comparison that should be made here is with the results achieved from the native SQL Server
striped dumps (see Chart-6), which can be likened to the ARCserve RAID-0 array backups. Backup
operations to ARCserve RAID-0 tape arrays outperform dump operations to similarly sized SQL
Server tape stripe sets, especially as more tape drives are added. With 8 DLT drives, ARCserve can
backup a SQL Server database almost as fast as SQL Server can with 11 DLT drives (about 35 GB/hr).
The tape block size used by both applications is the same (64KB).
This performance variance is interesting to note, especially since the ARCserve database backup agent
itself makes use of a SQL Server striped dump operation. The difference in the two procedures
however, is that with a SQL Server native striped dump, each thread reads data pages from disk into a
memory buffer, and from there writes the data out to each of the tape drives. With an ARCserve RAID
dump, each of the dump threads writes the data to a local named pipe created by the backup agent.
The backup agent provides a thread for each pipe to retrieve the data and send it to the ARCserve job
process, where it is written to the tape drives. Apparently the use of the named pipes mechanism
allows the ARCserve backup agent to process the incoming data and send it out to tape faster than the
native SQL Server process can. However, this mechanism also demands much more CPU power; with
¤
1997 Compaq Computer Corporation, All Rights Reserved
Page 45
Doc No 444A/0797