Testing Methodology

Basically the same as the SD2008 performance review, using 8 Gigabit ports of the Agilent N2X network tester. The ports were wired to switch ports 1, 10, 3, 12, 5, 14, 7, 16 in order to maximize the amount of traffic accross the various paths, if there's such a thing on this unit...

Various packets size from 46 to 1500 bytes (8 ports)

Packets size	Tx/Rx (pps) per port	Tx/Rx (pps) combined	Throughput (Mbps) per port	Throughput (Mbps) combined	Latency (µs)
46	1,488,095	23,809,529	761.90	12,190.48	4.11
64	1,225,490	19,607,844	803.92	12,862.74	4.59
128	753,012	12,048,193	879.52	14,072.29	5.40
256	425,170	6,802,720	931.97	14,911.56	6.58
512	227,273	3,636,366	963.64	15,418.19	9.62
1000	120,195	1,923,120	978.87	15,661.92	15.04 or 83.91(*1)
1500	81,275	1,300,399	987.00	15,792.05	21.47

*1 Surprise here: the switch can't do line-rate with 1000-bytes packets size ! It does 99% with no sweat, with a reasonable delay of 15µs. But at 100% line rate, the switch starts queueing and dropping packets: latency raises above 83µs and packets loss is 229 pps per port. Increasing the frame rate to exactly 120,195 pps on the Agilent, no packet loss. At 120,196 pps, I can see the delay building-up (RX or TX queue filling-up?). The switch is dropping 22 packets over a 30s run. To the switch fabric's credit, the loss is independant of the status of the other ports: with 7 oversubscribed ports out of the 8 I'm using, the 8th is showing optimal performance and low delay.

Wanting to know more about the 1000-bytes packets problem, I setup the tester to try out packets sizes between 990-1010 bytes. Some sizes exhibit packets loss and others don't ! Packet loss occurs for 999, 1000, 1001 bytes ip packets (add 18 for Ethernet frames). I didn't have the time to test this more thoroughly, but in practice this really is a non-issue.

Jumbo Frames, or Lack Thereof: Keep It to the Bare Minimum

The SR2016 fares even worse than my respectable SD2008: the switch passes ip packets up to 1500 bytes in size (that's 1518 bytes Ethernet frames), but increase this by just 1 octet and the switch stops passing any traffic. So forget about any fancy L2 encapsulation of your L3 traffic here, because the SR2016 does't play very nicely mtu-wise.

Last words

Pros:

• Performance is very good indeed - using 8 ports, I could prove that the fabric is very efficient, able to forward traffic at line rate
• Medium-sized form factor, stylish metal casing
• Internal, universal power supply

Cons:

• Noisy !
• MTU is limited to the standard ipv4/ethernet packets size and will be a turn-off for demanding users - no jumbo frames

<< Back to Part 1: Opening the Switch