Cisco Nexus 5548P vs. Arista 7148SX, IBM BNT G8264 and Juniper EX4500 competitive lab test

Brad Reese speaks out
Archive

Aironet

Competitive Lab Tests

Subscribe to Brad Reese speaks out

Cisco Nexus 5548P vs. Arista 7148SX, IBM BNT G8264 and Juniper EX4500 competitive lab test
Mon, 3/28/11 - 11:59pm View comments

IBM commissioned the independent test lab, The Tolly Group, to perform a competitive lab test that benchmarked 10GbE switches for price/performance, throughput, latency, microburst buffer capacity and power consumption.

The lab test pitted the IBM BNT G8264 RackSwitch against the following 3 competitive 10GbE switches:

1. Arista 7148SX

2. Cisco Nexus 5548P

3. Juniper EX4500

Tolly's price/performance results:

The price/performance ratio of the DUTs (device under test) were calculated in terms of US dollars per Gbps (US$/Gbps), using the maximum throughput as measured by Tolly's engineers. For the comparison below, Tolly configured the DUTs as similarly as possible:

According to Tolly, a 10GbE Top-of-Rack Data Center switch should perform at line rate across all frame sizes. Forwarding anything less than 100% of frames when required can cause retransmission of data, potentially resulting in degraded application performance.

Tolly engineers benchmarked the DUTs (device under test) for throughput and latency while handling bidirectional Layer-2 traffic from 64 to 9216 bytes in size as defined by RFC 2544.

For Tolly's ATSS port-to-port latency test, the Arista, Cisco and Juniper DUTs (device under test) were tested for the cut-through latency using 48 10GbE ports connected in port-pairs, from the outside ports inward, topology handling a load of 100% of the 10GbE line rate for Cisco and IBM. Other devices were tested at their maximum achievable rate for each DUT based on the IxAutomate ATSS test suite.

Comparison of Layer 2 throughput and latency results:

Tolly believes that depending on the type of protocol, as in multicast applications, frames can be dropped if line rate throughput is not achieved.

The Tolly Layer 2 multicast performance RFC 3918 One-to-Many throughput test evaluated the ability of the DUTs to forward multicast traffic from one transmitting 10GbE port to the remaining 47 10 GbE ports in a 1-to-47 "fan-out" configuration with 10 Gbps of Layer 2 unidirectional multicast traffic consisting of the standard Ethernet frame sizes between 64 bytes and 9216 bytes.

Comparison of Layer 2 multicast latency results:

Similar to the Layer 2 performance tests above, the switches were tested by Tolly for Layer 3 performance. To demonstrate the Layer 3 forwarding capabilities each port was configured in a separate subnet and a separate VLAN to require Layer 3 routing between ports.

For Tolly's ATSS port-to-port latency test the DUTs were tested for Layer 3 cut-through latency at their highest sustained throughput with the port-pairs connected as per the Layer 2 testing. Each port was configured in a separate subnet and a separate VLAN to require Layer 3 routing between ports.

Comparison of Layer 3 throughput and latency results (the Cisco Nexus 5548P did not support Layer 3 forwarding and was not tested):

According to Tolly, Microbursts are defined as sub-second periods of time when major bursts of network useage occur, causing utilization of network interfaces to become temporarily oversubscribed. This can possibly result in packet loss depending on the network device's capacity to buffer the excess packets.

These conditions are typically caused when many devices transmit to a single device at the same time; the Microburst can even occur in a period of nanoseconds and result in dropped packets. Detection of such instances may also be nearly impossible to indentify without the use of specialized network monitoring tools.

The below Microburst tests were run at the maximum achieveable throughput rate of each packet size for each DUT (device under test) using only 3 ports:

Tolly engineers benchmarked the power consumption by using a Watts up? PRO power meter to measure the power consumption of all DUTs (device under test) in similarly-equipped configurations while idle and while handling 50% and 100% line-rate bidirectional, full-mesh traffic of 64, 512, and 1518 byte frames.

At each load, traffic was maintained for 60 seconds and the average power consumption recorded. Tests were repeated 3 times and the results averaged. Since all the switches were equipped with dual power supplies, Tolly engineers utilized a 1 to 2 breakout cable, which aggregated the power draw to the power meter: