Compaq ProLiant 6000 Compaq DLT Tape Array II: High-Performance Backup of Ente - Page 5

Processor Quantity, Memory, PCI Bus

Page 5 highlights

ECG075.0997 (cont.) DLT Tape Array II ... HIGH-SPEED BACKUP RESEARCH Single and multi-processor systems were used to identify hardware factors that affect backup performance, including processors, memory, SCSI, and PCI channel bottlenecks. Some tests were run in a network environment; however, the focus was on local backups, because network bottlenecks significantly reduce feed speed. Future testing will address this problem and explore other ways to improve network-based backups. Listed below are test machine configuration details, along with factors that were taken into consideration in developing the highest performance configurations. Processor Quantity The average backup uses from 6 to 25 percent of a single 200 MHz Pentium Pro (256K cache version) processor's power. The lower number is typical when using a single disk or tape drive; usage percentage grows substantially when striping data across a redundant array of independent tape drives (RAIT) in a RAIT-5 configuration. All testing was performed on Compaq ProLiant 5000 and ProLiant 6000 systems so that up to four 200 MHz Pentium Pro processors could be used. An assumption was made that most large systems would have at least two processors, and the majority of tests were performed with that number. Memory A system with 256 MB of RAM (regardless of the number of processors) was the standard for all tests, to eliminate the remote possibility of complications due to inadequate RAM. PCI Bus Although some testing was done with EISA bus-based systems, the higher throughput of PCI (up to three times faster) clinched the decision to use it for high-speed backup testing. The 32bit x 33 MHz PCI bus peaked at 454 GB/hr in load tests with six Compaq SMART-2/P Array Controllers. To maximize bandwidth, tests were also performed to determine optimum card placement in the dual PCI bus system of the ProLiant 5000 and ProLiant 6000. The dual peerto-peer PCI buses connect to a common 1,919-GB/hr, or 540megabytes per second (MB/s) Gunning Transceiver Logic (GTL+) processor bus (the host bus), which permits a path to the processor that is twice the width of buses in many servers on the market today (Figure 4). Host Bus 540 MB/s Pentium Pro Processors HostBridge to-PCI Bridge Memory Slots 133 MB/s Secondary PCI Bus 133 MB/s Primary PCI Bus PCI-to-EISA Bridge Slots 33 MB/s EISA Bus Slots Figure 4: ProLiant 5000 peer-to-peer bus architecture. The ProLiant 5000 has seven PCI slots and three EISA slots across the two PCI buses (eight total; two are shared). The ProLiant 6000 has eleven slots: nine PCI and two EISA.

  • 1
  • 2
  • 3
  • 4
  • 5
  • 6
  • 7
  • 8
  • 9
  • 10
  • 11
  • 12
  • 13
  • 14
  • 15
  • 16
  • 17
  • 18
  • 19
  • 20
  • 21
  • 22
  • 23
  • 24
  • 25
  • 26
  • 27
  • 28
  • 29
  • 30
  • 31
  • 32
  • 33
  • 34
  • 35
  • 36

7
%
(cont.)
DLT Tape Array II
ECG075.0997
H
IGH
-S
PEED
B
ACKUP
R
ESEARCH
Single and multi-processor systems were used to identify hardware factors that affect backup
performance, including processors, memory, SCSI, and PCI channel bottlenecks. Some tests
were run in a network environment; however, the focus was on local backups, because network
bottlenecks significantly reduce feed speed. Future testing will address this problem and
explore other ways to improve network-based backups.
Listed below are test machine configuration details, along with factors that were taken into
consideration in developing the highest performance configurations.
Processor Quantity
The average backup uses from 6 to 25 percent of a single 200 MHz Pentium Pro (256K cache
version) processor’s power. The lower number is typical when using a single disk or tape drive;
usage percentage grows substantially when striping data across a redundant array of
independent tape drives (RAIT) in a RAIT-5 configuration. All testing was performed on
Compaq ProLiant 5000 and ProLiant 6000 systems so that up to four 200 MHz Pentium Pro
processors could be used. An assumption was made that most large systems would have at least
two processors, and the majority of tests were performed with that number.
Memory
A system with 256 MB of RAM (regardless of the number of processors) was the standard for
all tests, to eliminate the remote possibility of complications due to inadequate RAM.
PCI Bus
Although some testing was done with EISA bus-based systems, the higher throughput of PCI
(up to three times faster) clinched the decision to use it for high-speed backup testing. The 32-
bit x 33 MHz PCI bus peaked at 454 GB/hr in load tests with six Compaq SMART-2/P Array
Controllers. To maximize bandwidth, tests were also performed to determine optimum card
placement in the dual PCI bus system of the ProLiant 5000 and ProLiant 6000. The dual peer-
to-peer PCI buses connect to a common 1,919-GB/hr, or 540megabytes per second (MB/s)
Gunning Transceiver Logic (GTL+) processor bus (the host bus), which permits a path to the
processor that is twice the width of buses in many servers on the market today (Figure 4).
Slots
Pentium Pro
Pentium Pro
Processors
Processors
Host Bus
Host Bus
Memory
Memory
Secondary PCI Bus
Secondary PCI Bus
Host-
Host-
to-PCI
to-PCI
PCI-to-EISA
PCI-to-EISA
EISA Bus
EISA Bus
33 MB/s
133 MB/s
133 MB/s
Primary PCI Bus
Primary PCI Bus
540 MB/s
Bridge
Bridge
Bridge
Slots
Slots
Figure 4: ProLiant 5000 peer-to-peer bus architecture. The ProLiant 5000 has seven PCI slots
and three EISA slots across the two PCI buses (eight total; two are shared).
The ProLiant 6000 has eleven slots: nine PCI and two EISA.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.