A6826A – PCI-X Dual port 2Gb/s Fibre Channel Adapter
Performance Paper for PCI platforms
Table of contents
Introduction...............................................................................................................................................2
Executive summary.....................................................................................................................................2
Test results ................................................................................................................................................3
IOPs .....................................................................................................................................................3
Service demand.....................................................................................................................................3
Throughput............................................................................................................................................4
Scalability .............................................................................................................................................5
System configuration guidelines...................................................................................................................7
Test details................................................................................................................................................8
Products used in testing...........................................................................................................................8
Test configuration...................................................................................................................................9
Additional Information................................................................................................................................9
1
Introduction
This paper provides basic I/O performance and scalability information for the A6826A dual-channel 2Gb Fibre
Channel adapter on HP’s PCI platforms. A series of tests were conducted to evaluate the adapter’s performance
on HP servers, namely rp5450, rp7400, rp8400 and Superdome. The tests measured the IOPS and throughput of
single and multiple adapters.
This paper addresses the performance capabilities of A6826A when used with the above mentioned HP
platforms. The paper also discusses the system setup considerations to obtain the maximum possible performance
from A6826A on these platforms.
This paper focuses on the following topics:
• Test results: Single and multiple adapters performance data, which include the IOPS and throughput
for single and dual ports, will be discussed.
• Scalability: Multiple adapter scalability tests on systems namely Superdome, rp8400, rp7400 and
rp5450 will be discussed.
• System configuration guidelines: Superdome, rp8400, rp7400 and rp5450 system configurations
and recommendations will be discussed.
• Test details: HP’s products used, test setup, benchmark tool used and system configuration in the test
setup will be discussed.
Executive summary
A6826A on HP mid range and high end platforms offers an excellent SAN solution. A6826A provides read
throughput of 195MB/s, and write throughput of 188MB/s on a single port. With dual ports a read throughput of
390MB/s, and write throughput of 368MB/s are achieved.
A6826A performance summary
Operation
iops 35530 21300 20300 70100 42700 31300
throughput
(MB/s)
NOTE:
*bd: Bidirectional Operation.
**The full duplex throughput is limited by 66MHz PCI bandwidth (528 MB/s).
The above performance data is obtained on A6826A seated in a 66MHz PCI slot of a single cell single CPU
(750MHz) rp8400.
achieved 195 188 351 390 368 412**
% of 2Gb/s FC
theoretical max
read write bd* read write bd
single port two ports
97.5 94 87.7 97.5 92 51.1
On rp8400 and Superdome A6826A provides outstanding performance with linear scaling up to 4 cards, when
all the eight ports are being used at 2Gb/s bandwidth. Additional cards may be added to provide greater
connectivity. The card delivers excellent performance in limited configurations on rp7400 and rp5450.
2
Test results
Diskbench (db) utility is used to generate the read and write test traffic. Various block size tests were executed for
read and write on single and dual ports. The IOPS metric are obtained with 1 KB block size transfers. The
throughput metric is obtained with 128 KB block size transfers. The throughput metric is useful in modeling large
sequential transfers such as remote backup etc. The IOPS metric is useful in modeling small transactional traffic.
IOPs
Chart 1a: IOPS (in thousands) Chart 1b: % CPU Utilization
IOPs
Read
Write
80000
70000
60000
50000
40000
30000
20000
10000
0
12
35530 70100
21300 42700
Numbe r of P orts
% CPU Uitlization
Read
Write
100
90
80
70
60
50
40
30
20
10
0
12
40.02 92.85
25.77 54.29
Number o f Po rts
The chart, chart 1a, shows the number of IO operations per second for read and write operations on single and
dual ports of A6826A. The X-axis is the number of ports and the Y-axis the number of IO operations per second.
The chart 1a shows linear scaling of IOPS for single and dual ports. The number of read IO operations per
second for single port is 35530 and for dual ports are 70100. The number of write operations per second for
single port is 21300 and for dual port are 42700. The IOPS metric for A6826A is limited by the processor used
on A6826A.
The chart 1b shows the CPU utilization for the IO operation tests.
The chart 1b shows the % CPU utilization for IO operation tests. The X-axis is number of the Fibre Channel ports.
The Y-axis is the % CPU utilization. Only one 1 CPU is configured in these tests, hence this is a single CPU
utilization.
Service demand
A6826A offers a great service demand for small size IO operations. To illustrate service demand of A6826A
small size operations tests were conducted using db. Read and write operations of IO sizes of 4KB and 8KB were
performed on a single cell single CPU rp8400 and the results were recorded.
The following table shows single and dual port throughput and CPU utilization for Sequential Read and Sequential
Writes with 4KB and 8KB IO sizes.
Number
of Ports
Thru
1
CPU2
Single 130 37%
Dual 275 74%
NOTE:
1
Throughput in MB/s
2
Single CPU utilization
3
Service Demand =
Data gathered on single cell single CPU rp8400
Sequential Read Sequential Write
4KB 8KB 4KB 8KB
SD3
Thru CPU
2.93
2.75
((%CPU utilization/100)/Throughput in KB/sec)
190 25%
379 65%
SD
1.34
1.75
Thru CPU
83 25%
164 53%
SD
3.08
3.30
Thru CPU
127 20%
248 58%
SD
1.61
2.39
3
Throughput
Chart 2a : Single and Dual Port Throughput Chart 2b : CPU Utilization for Throughput tests
450
2Gb/s FC
Therotical Max
Read
Write
Throughput (MB/s)
400
350
300
250
200
150
100
50
0
12
200 400
195.15 390.03
188.13 368.77
Number of Ports
Data obtained on a single cell, single CPU, rp8400
14
12
10
8
6
4
2
%CPU utilization
0
12
Read
Write
6.53 11.54
6.21 12.01
Number of Ports
The chart 2a shows the read and write throughput of single and dual ports of A6826A. The X-axis is the number
of Fibre Channel ports. The Y-axis is the throughput in MB/s.
As the chart 2a shows, one port of A6828A performs read at 195MB/s. Two ports scale at 2x the performance
of single port performing reads up to 383MB/s.
For the write operations, one port of A6826A performs write operations at 188MB/s. Two ports scale
approximately at 2x the performance of the single port performing writes up to 368MB/s.
The chart 2a demonstrates outstanding read and write performance with linear scaling for 1 and 2 ports of
A6826A.
The chart 2b shows the % CPU utilization of throughput tests. The X-axis is the number of the Fibre Channel port.
The Y-axis is the % CPU utilization. Only one 1 CPU is configured in these tests, hence this is a single CPU
utilization. The charts 2b shows very small percentages of CPU utilization for single and dual ports in read and
write throughput tests.
4
Scalability
A series of tests were conducted to evaluate the scalability aspect of A6826A on various HP systems, namely
Superdome, rp8400, rp7400 and rp5450.
Chart 3a: 128KB Sequential Read
1800
1600
1400
1200
1000
800
600
400
200
Throughput (MB/s)
0
12345678
Superdome
rp8400
rp7400
192 383 572 766 956 1147 1336 1524
191 382 572 762 954 1144 1334 1523
191 381 571 712 710 899 1001 1026
Number of Por ts
2 Cell Superdome Partition with 8 (850MHz) CPUs
2 Cell rp8400 Partition with 8 (750 MHz) CPUs
rp7400 with 8 (550 MHz) CPUs
The above chart, chart 3a, shows the read throughput performance of 1, 2, 3 and 4 A6826A in Superdome,
rp8400 and rp7400. The X-axis is the number of ports. Each port represents one port of A6826A. The Y-axis is
the throughput performance in MB/s.
The chart shows excellent read throughput performance with linear scaling for 4 A6826As on Superdome and
rp8400 reaching an aggregate of 1524MB/s. Superdome and rp8400 along with A6826A offer a great
scalable and connectivity SAN solution.
On rp7400 the read throughput scales linearly up to 4 ports (2 A6826As) and reaches a plateau from 4 to 5
ports and then ascends up to 1025MB/s for 8 ports. The decline in throughput from 7 to 8 results from the IO
subsystem limitations on rp7400. rp7400 offers a good scalable and connectivity SAN solution.
5
1600
1400
1200
1000
800
600
400
200
Throughput (MB/s)
Chart 3b: 128 KB Sequential Write
0
12345678
Superdome
rp8400
rp7400
188.35 318.93 494.53 600.46 783.03 906.7 1073.9 1183
188.21 327.57 507.1 638.67 794.84 882.5 1048.3 1139.1
188.6 375.87 557.2 722.77 909.45 1077.6 1255.8 1413.7
Number of Ports
2 Cell Superdome Partition with 8 (850MHz) CPUs
2Cell rp8400 Partition with 8 (750MHz) CPUs
rp7400 with 8 (550 MHz) CPUs
The above chart, chart 3b, shows the write throughput performance of 1, 2, 3 and 4 A6826A in Superdome,
rp8400, rp7400 and rp5450 systems. The X-axis is the number of ports. Each port represents one port of
A6826A. The Y-axis is the throughput performance in MB/s.
The chart shows excellent write throughput performance with linear scaling for 4 A6826As on Superdome and
rp8400 reaching an aggregate of 1180 MB/s. Superdome and rp8400 along with A6826A offer a great
scalable and connectivity solution.
On rp7400 the write throughput scales linearly for 4 A6286As reaching an aggregate of 1413MB/s.
The table below summarizes the maximum number of cards that can be used on these platforms in order achieve
linear scalability. Additional cards can be used to provide better connectivity.
Maximum number of A6826A for linear scalability Ports operating
speed
Both Ports at 2Gb 4 2 per IO Cage 4 2
One port at 1 Gb and
other at 2Gb
Both ports at 1 Gb 8 4 per IO Cage 8 4
rp8400 Superdome rp7400 rp5450
6 3 per IO Cage 6 4
6
System configuration guidelines
Superdome
To get the throughput and scalability with multiple cards on a Superdome, HP recommends the following:
The cells should be configured with a multiple of 8 equal capacity DIMMs to take advantage of memory
interleaving. The 8 DIMMs should be evenly distributed across the two busses.
Each IO cage of Superdome offers twelve PCI slots, eight 2X and four 4X PCI slots. The 2X slots have 265MB/s
bandwidth and 4X PCI slots have 530 MB/s bandwidth. HP recommends A6826A to be installed in a 4X PCI slot
(Physical slot number 4, 5, 6 or 7) in order to achieve performance shown in this paper. To get a linear scalability
HP recommends up to a maximum of two A6826As per IO cage.
rp8400
To get the throughput and scalability with multiple cards on an rp8400, HP recommends the following:
The cells should be configured with a multiple of 8 equal capacity DIMMs to take advantage of memory
interleaving. The 8 DIMMs should be evenly distributed across the two busses.
rp8400’s IO subsystem offers sixteen PCI slots, two 2X PCI and fourteen 4X PCI slots. The 2X slots have 265MB/s
bandwidth and 4X PCI slots have 530 MB/s bandwidth. HP recommends A6826A to be installed in a 4X PCI slot
(Physical slot numbers 1, 2, 3, 4, 5, 6) to achieve performance shown in this paper. To get a linear scalability HP
recommends a rp8400 should have at least 2 cell partition and with two A6826As installed in any of 1 to 8 4X
PCI slots and a maximum of two A6826As in any of the 9 to 16 4X PCI slots.
rp7400
To get the throughput and scalability with multiple cards on an rp7400, HP recommends the following:
The rp7400 supports up to 4 memory carriers each providing 8 memory slots. All the carriers should be
populated with equal size DIMMs to maximize the memory bandwidth.
rp7400 offers twelve PCI slots, two 2X (a.k.a. Turbo) and ten 4X (a.k.a. Twin Turbo) PCI slots. HP recommends
A6826A to be installed in a 4X slot (Physical slot numbers 3, 4, 5, 6, 7, 8, 9, 10, 11, 12) to achieve
performance shown in this paper. To achieve a linear scalability HP recommends up to a maximum of two
A6826As to be installed in the right IO backplane (Physical slots 7 to 12) and up to a maximum of two A6826As
to be installed in the left IO backplane (Physical slots 1 to 6).
rp5450
rp5450 is designed to be low end mid range system with limited IO capabilities. HP recommends A6826A on
rp5450 to be used as a connectivity solution.
7
Test details
The performance results presented in this paper were obtained with A6826A on various HP platforms. The system
configurations in the test setup are tabulated in the following table:
Products used in testing
products used diskbench for performance measurement test
Servers
Tested
Benchmark software
rp8400
• 2 cell partition
• Each cell with
2 - 750MHz PA-8700
CPUs
4GB System Memory
• HP-UX B.11.11 OS
• 2Port 2Gig Fibre Channel driver
(Ver. B.11.11.01)
Superdome
• 2 Cell partition
• Each cell with
2 - 875 MHz PA-? CPUs
4 GB System Memory
• HP-UX B.11.11 OS
• 2Port 2Gig Fibre Channel driver
(Ver. B.11.11.01)
rp7400
• 8 – 550 MHz PA8600 CPUs
• 8GB System Memory
• HP-UX B.11.11 OS
• 2Port 2Gig Fibre Channel driver
(Ver. B.11.11.01)
rp5450
• 4- 440 MHz PA8500 CPUs
• 4GB System Memory
• HP-UX b.11.11 OS
• 2Port 2Gig Fibre Channel driver
(Ver. B.11.11.01)
A6826A HBA
• PCI-X dual–channel 2Gb Fibre
Channel adapter
• Each port capable of
independently operating at 1 or
2 Gb/s (Auto-Negotiation)
• 33/66/100/133MHz-64bit
Diskbench (db) is the benchmark suite
db
that generated disk read and write traffic
for these tests.
HP StorageWorks 2Gb/s Disk System
DS2405
8
Test configuration
The test configuration consists of 8-way (750 MHz) rp8400 with 8 GB of System memory and four A6826A HBAs.
Two HBAs were connected to HP 2Gb/s Fibre Channel switches and the other two were connected to another HP
2Gb/s Fibre Channel switch. Sixteen HP DS2405 were evenly connected to the two switches. The switches were
zoned so that each port of A6828A could see 2 DS2405.
Similar test setup was used to collect performance data on Superdome, rp8700 and rp5450.
Single CPU system configuration was to obtain IOPS and throughput results. Additional CPUs were added to study the
scalability of A6826A on rp7400, rp8400 and Superdome.
test setup
Additional Information
For more information about A6826A please visit
http://www.hp.com/products1/unixserverconnectivity/storagesnf2/index.html
© 2003 Hewlett-Packard Development Company, L.P. The information
contained herein is subject to change without notice. The only warranties for
HP products and services are set forth in the express warranty statements
accompanying such products and services. Nothing herein should be construed
as constituting an additional warranty. HP shall not be liable for technical or
editorial errors or omissions contained herein.
Intel and Itanium are trademarks or registered trademarks of Intel Corporation
in the U.S. and other countries and are used under license. Pentium is a U.S.
registered trademark of Intel Corporation. Microsoft, Windows, and Windows
NT are U.S. registered trademark of Microsoft Corporation. UNIX is a
registered trademark of The Open Group.
07/2003
5981-7419EN
.
9
Loading...