We assessed the Arista DCS-7508 with tests of performance and power consumption. The performance tests used the Spirent TestCenter test instrument to measure layer-2 and layer-3 unicast throughput and latency; layer-2 and layer-3 multicast throughput and latency; OSPF equal cost multipath (ECMP); N+1 fabric failover; buffering capacity; OSPF routing capacity; and multi-chassis link aggregation (MLAG). All tests used version 4.8.4 of Arista's EOS software and Version 3.95 of Spirent TestCenter software.
The unicast performance tests involved a fully meshed traffic pattern among 384 10G Ethernet switch ports. The Spirent TestCenter generator/analyzer offered traffic for a duration of 60 seconds, and measured the latency of every frame received. We repeated this test with 64-, 256-, 1,518-, and 9,216-byte Ethernet frames.
The layer-2 multicast traffic tests involved a traffic pattern with one transmitter port and 383 receiver (subscriber) ports. Here, all 383 receiver ports on the Spirent TestCenter instrument joined the same 4,095 multicast groups using IGMPv3. After the switch's IGMP snooping table was fully populated, the test instrument then offered traffic to the single transmit port, with destination addresses of all 4,095 multicast groups. As in the unicast tests, the instrument measured throughput and latency for four frame lengths.
The layer-3 multicast tests used the same traffic pattern as the layer-2 tests, with one transmitter port and 383 receiver (subscriber) ports. In this case, however, all 384 switch ports also ran the protocol independent multicast-sparse mode (PIM-SM) routing protocol, as did the Spirent TestCenter test instrument. All switch ports used PIM-SM to learn multicast routes. Then, all 383 receiver ports on the Spirent TestCenter instrument joined the same 512 multicast groups using IGMPv3. The instrument measured throughput and latency for the same four frame lengths as in the other performance tests.
We verified 16-way ECMP functionality by bringing up 16 OSPF adjacencies using the Spirent test instrument, each advertising networks using the same metric. An additional 16 Spirent ports emulating hosts then exchanged traffic across the Arista switch with the routes learned using OSPF. We then inspected outbound packet counts on the 7508's OSPF interfaces to verify the device load-balanced traffic across all links.
To measure fabric failover time, we configured the system to use layer-2 unicast switching. While the Spirent test instrument offered 64-byte frames at wire speed in a fully meshed pattern to all 384 ports, we physically removed the switch's primary supervisor module, forcing the system to bring up the secondary supervisor module. At the end of the test, we derived failover time by dividing frame loss into offered frame rate.
We measured buffering capacity with three sets of tests. The first of these offered traffic from 256 ingress ports to 128 egress ports using a partially meshed pattern as described in RFCs 2285 and 2889. In the first test, we offered traffic for 30 seconds and determined whether frame loss was over or under the expected amount of 50% (given the 2:1 overload).
In the second buffering test, we configured 128 sets of ports so that two ports sent traffic to one port - again creating a 2:1 overload, but in a non-meshed pattern. All ingress ports were on different line cards than all egress ports. Here, we offered various lengths of bursts (for example, 50,000 frames to each of two ingress ports) to determine the longest burst possible without frame loss.
In the third buffering test, we offered bursts of traffic to 383 ingress ports, all destined to one egress port. Again, this test used bursts of varying lengths to determine the largest amount of traffic the switch could buffer without loss.
To measure OSPF routing capacity, we configured the Spirent instrument to act as an OSPF neighbor on one port of the Arista system, and as an emulated host on another port. We offered varying numbers of summary link-state advertisements and then offered traffic from the "host" to determine if the Arista switch could forward traffic to all networks advertised. If a test failed, we restarted the OSPF process and tried again with a smaller number of routes. We repeated using a binary search until finding the maximum number of routes the system could handle.
We measured MLAG functionality and failover time. Both tests used four switches: A pair of 7508s acted as MLAG peers, and a pair of 7050 switches each had eight-way MLAG connections on each side of the MLAG peers (each 7050 had four attachments to each 7508). We verified MLAG functionality by attaching the Spirent instrument to the 7050s and offering bidirectional traffic for 30 seconds from 256 emulated hosts across the MLAG network. We then inspected outbound packet counts on the 7508's MLAG interfaces to verify the device load-balanced traffic across all links.
To measure MLAG resiliency, we repeated the functionality test for 300 seconds, and cut power to one 7508 during the test. As in the fabric failover tests, we derived recovery time by dividing frame loss into offered frame rate.
We measured power consumption in fully loaded and 50% loaded cases. In the fully loaded case, all 384 10G Ethernet ports forwarded traffic at wire speed. In the 50% loaded case, we removed four of the system's eight line cards, and offered traffic at wire speed to all remaining ports. We used a Fluke 335 clamp meter to measure voltage at the power outlet and amperage used at each of two power supplies; we multiplied these measurements to obtain watts.
Read more about lan and wan in Network World's LAN & WAN section.