APA doesn't perform as expected

[Michael Peller]

Currently I'm evaluating the benefits of the APA software. I set up the following dedicated test szenario: 2 N4000-44 each with 2 Gigabit Adapters A4926A-SX. 2 Cisco C4006 switches. Each N connects to its dedicated switch. The switches are connected together using an uplink channel built from 2 Gigabit lines. The APA software uses LB_MAC load balancing, the ports are using FEC_AUTO. Each switch and its N set up the link aggregate properly (output of 'show port channel' and lanadmin). Running synthetic benchmarks using the above config I'm getting an througput of 86 MB/s (avg). In contrast, running without APA and using only one Gigabit adapter the throughput is 76 MB/s (avg). Using 2 adapters with separate LANs I got 120 MB/s (avg)(Btw, it seems, that I'm CPU limited here). I made the tests with LB_CPU, too. But the results are similar to LB_MAC. Btw, running the benchmarks using the loopback device the N delivers 280 MB/s (avg). Now, my question: is it possible to get more than 86 MB/s out of the link aggregate? Any hint is appreceated. At the moment dedicated LANs perform/scale much better...

[sung-hoon chang]

Hello

I think, you need some more servers for performance upgrade.

LB_MAC use mac address for load balancing so when a N-class have a network session, the session have only one mac address. Because the APA does not work for load balancing.

If you have some more servers for the benchmark, when a N-class have at least several network session so have several mac address and the load balancing will be worked. And then your performance of APA network will be upgraded.

And LB_CPU is almost much the same with case of LB_MAC.