Chart 6 - Striped Dumps To 15/30 Dlt Tape Drives - Compaq 219700-001 - ProLiant - 1500 White Paper

Compaq Backup and Recovery for Microsoft SQL Server 6.x Page 40

Chart 6 - Striped Dumps to 15/30 DLT Tape Drives

Starting with 1 drive yields the 5.00 GB/hr throughput seen before with a single drive. As Chart-6
shows, the expected increase in throughput is then seen as additional drives are added to the stripe set.
Looking at the chart, we notice a general trend of "diminishing returns" whereby the gain in
throughput achieved by adding drives decreases successively as more drives are added. Earlier, we
predicted that the ability of the Smart-2 controller to deliver data would restrict the system to under 50
GB/hr. However, this trend of decreasing gains actually begins much earlier, and continues so that the
system never reaches the 50 GB/hr maximum. From 1 to 8 drives our scalablility (performance
increase) is actually quite good - an average of 3.5 GB/hr increase for each additional drive. From 9 to
12 drives the scalability is more modest - averaging to a 2.0 GB/hr increase per additional drive. We
therefore recommend that no more than 8 15/30-GB (or 10/20-GB) DLT drives be used in a single
system for this type of backup activity. Production environments that wish to use the SQL Server
dump but that require a backup rate 30GB/hr or greater to meet their time constraints, should consider
a solution using either disk drives (previous section) or 35/70-GB DLT drives (next section).
The reason for the non-linear scalability probably pertains to a combination of hardware and
application behavior. Ideal scalability in complex systems is rarely achieved, and in this case data
traveling from one media (disk) to another media (tape) has to go through numerous intermediate
hardware paths (SCSI bus, PCI bus, system memory, etc.). Although we can see from the above chart
that neither the system processors nor the PCI bus is close to saturation, there are still latencies
associated between each of the hardware paths which can combine to cause overall system scalability
to fall-off as load increases. Furthermore, we should remember that although multiple threads of
operation are being used by SQL Server to deliver data to multiple tape drives, all of the threads are
reading from a single source (one database) and are forced to alternate between each other so that a
thread which has requested data for its tape drive cannot again submit a read request until all other
threads have done so. This 'round-robin' processing among the different device threads brings about a
certain amount of serialization in the application that affects scalability.
Note that when doing a SQL Server striped dump it is possible to combine different tape drives, or
even different types of dump devices, into a stripe set. We do not recommend however, that DLT