HP Auto Port Aggregation Performance and Scalability

Bidirectional Throughput Performance

Figure 2 shows HP APA throughput numbers with bidirectional transmit and receive workloads for 2-, 4-,

and 8-port link aggregates.

Figure 2 HP APA Throughput for Bidirectional Transmit and Receive Workloads

2000

4000

6000

8000

10000

12000

14000

16000

Bidirectional throughput

3682 7320 13524

2-port APA 4-port APA 8-port APA

Megabits per Second

Figure 2 shows that HP APA provides an exceptional sustained data rate of 3.68 Gb/s for a 2-port aggregate

and 7.3 Gb/s for a 4-port aggregate, which is 97 percent of link rate. The 8-port aggregate achieved an

excellent sustained data rate of 13.5 Gb/s, which is close to 90 percent of link rate.

For these tests, the server CPU utilization (normalized across all 16 CPUs on the server) was approximately

20 percent, 41 percent, and 92 percent for 2-port, 4-port, and 8-port link aggregates, respectively. The

average packet rate was around 154K packets per second per port.

Configuration

For the best performance and scalability with HP APA, HP recommends the following:

• Use processor affinity for processes that generate networking traffic over the HP APA link aggregate.

For best results, bind a work process to the CPU that handles most of the inbound traffic for that process

as follows:

1. Find the APA link aggregate ports on the server that are used for processing the inbound traffic

for a process. The load balancing algorithm for the trunk on the switch determines this. For

information on the load balancing algorithms used for trunk configurations on the switch, see the

documentation for your switch.

2. Ensure that the inbound processing for all workloads is evenly distributed across all processors on

the server. Be sure that no two ports have the same CPU doing all the inbound processing for both

ports. Ideally, assign all ports unique interrupt CPUs. If necessary, change the interrupt CPU

configuration for ports from the default configuration. Use the intctl command to change the

interrupt CPU configuration. See intctl(1M) for more information.

3. Set the processor affinity when starting the process or change the processor affinity for an already

running process, by using the mpsched command. See mpsched(1M) for more information.

NOTE If the work processes are not explicitly configured for processor affinity, you might see a

noticeable increase in CPU utilization. In addition, you might also see a decrease in the aggregate

Configuration 9