Juniper EX 4200 Virtual Chassis Performance
Juniper EX 4200 Virtual Chassis Performance
Juniper EX 4200 Virtual Chassis Performance
May 2010
Executive Summary
Juniper Networks commissioned Network Test to assess the performance of Virtual Chassis technology on its EX Series Ethernet switches, especially in the areas of latency and high availability. The key objectives were to determine whether Virtual Chassis configurations would provide low latency and fast recovery from network failures. Among the main findings of this project: Latency is always lower between Juniper EX4200 switches when those switches use a Virtual Chassis configuration rather than a standalone configuration. Test results clearly show reduced delay, sometimes by better than a 2:1 margin, when using Virtual Chassis configurations. Virtual Chassis latency is lower in both Ethernet switching and OSPF routing configurations. Virtual Chassis configurations recover from hardware and software failures in far less than 1 second. Tests involved redundancy protocols such as rapid spanning tree and Juniper Redundant Trunk Group as well as various hardware failures. A Virtual Chassis configuration took just 6 microseconds to recover from loss of power to its master switch. Virtual Chassis interconnections operate at 30-Gbit/s rates in each direction between switches for all frame sizes. Channel capacity between Virtual Chassis member switches was never a bottleneck in this project.
Core/Aggregation Layer
L2/L3 boundary
EX8216 EX8216
Access Layer
L2 Access
M
East-West traffic Spirent TestCenter 10 x GbE ports for each EX4200 member switch Spirent TestCenter 80%
Figure 2: Standalone (non-Virtual Chassis) Test Bed Network Test offered traffic using the 80/20 pattern described above, and measured latency on the first four switches (those in the lower-left hand corner of Figures 1 and 2). Then engineers reconfigured switches in the Virtual Chassis configuration shown in Figure 1 above. In all, engineers repeated these tests three times to compare standalone and Virtual Chassis latency: Standalone with all ports in a single VLAN. This is a pure layer-2 configuration, as shown in Figure 2 above. Virtual Chassis configuration with all ports in a single VLAN. This is a pure layer-2 configuration, as shown in Figure 1 above. Virtual Chassis configuration with all ports in different VLANs/IP subnets and with OSPF running. While the physical topology is identical to the configuration shown in Figure 1, this is a combination layer-2/layer-3 setup. Here, the boundary between switching and routing is at the access layer; in contrast, that boundary is at the core/aggregation layer in the first two configurations. This final configuration also demonstrates the ability of Virtual Chassis technology to work equally well in layer-2 and layer-3 configurations. For each set of tests, engineers offered the same number of frames at line rate, and measured the latency of every single frame. This made it possible to make meaningful comparisons between the standalone and Virtual Chassis test cases.
Average latency (microseconds) Standalone, same switch L2 Virtual Chassis, same switch L3 Virtual Chassis, same switch Standalone, switch 0 <-> switch 1 L2 Virtual Chassis, switch 0 <-> switch 1 L3 Virtual Chassis, switch 0 <-> switch 1 Standalone, switch 0 <-> switch 2 L2 Virtual Chassis, switch 0 <-> switch 2 L3 Virtual Chassis, switch 0 <-> switch 2 Standalone, switch 0 <-> switch 3 L2 Virtual Chassis, switch 0 <-> switch 3 L3 Virtual Chassis, switch 0 <-> switch 3
Table 1: Standalone vs. Virtual Chassis Latency
Maximum latency (microseconds) 5.44 8.19 9.24 20.81 8.96 10.22 9.58 11.08 12.36 24.40 14.37 15.72
The most notable result is that latency between switches is always significantly lower in Virtual Chassis configurations, both for average and maximum delay. The reason for the latency improvement with a Virtual Chassis configuration is simple: Frames do not need to traverse the core/aggregation switches, and thus save the delay added by one or more extra hops through the network. For traffic within the same switch, latency is slightly higher in the Virtual Chassis test cases. Although both the standalone and Virtual Chassis test cases involve the same fully meshed traffic pattern among all switch ports, the workload is slightly heavier in the Virtual Chassis test case. In that instance, traffic between all ports on other Virtual Chassis member switches traversed each switch in the Virtual Chassis configuration, imposing a slightly heavier load. In the standalone case, there is no such background load: Each switch offloads all nonlocal traffic to the core/aggregation layer for forwarding, and thus only handles switching of local traffic. However, for any configuration involving multiple switches, average and maximum latency are lower often by a factor of 2 when using a Virtual Chassis configuration rather than a standalone configuration.
Core/Aggregation Layer
EX8216 EX8216
Access Layer
EX4200 Virtual Chassis 1
M B
Spirent TestCenter
Spirent TestCenter
Figure 5 below shows the topology used for the RE failure test. Here, engineers powered off the master switch in the Virtual Chassis configuration (labeled M in the figure), forcing the backup switch (labeled B) to take over. Virtual Chassis member switches also used RTG in both the VCP and RE failover tests. As in the protocol failure tests, engineers derived failover time from frame loss measured during the cutover period.
Core/Aggregation Layer
L2/L3 boundary
EX8216 EX8216
Access Layer
L2 Access
M
Spirent TestCenter
Test case Layer 2, rapid spanning tree (RSTP) failover Layer 2, redundant trunk group (RTG) failover Layer 2, Virtual Chassis port (VCP) failure Layer 3, VCP failure Layer 2, Routing Engine (RE) failure with RTG Layer 3, RE failure Layer 3, OSPF failover Table 2: Virtual Chassis Failover Times
Failover time (milliseconds) 98.520 96.770 0.059 0.116 0.006 0.213 285.451
In all test cases, the Virtual Chassis configuration recovered from failures in far less than 1 second. Even in the case of the longest failover time, for OSPF, it took less than 300 milliseconds for the Virtual Chassis configuration to recover from a link failure and forward traffic along a recomputed path. Recovery times were lower still for other protocols. The Virtual Chassis configuration converged following RSTP and RTG failures in less than 100 milliseconds. The fastest recoveries were those following hardware failures. In the fastest failover of all, it took just 6 microseconds to recover from the loss of power to a master switch in the Virtual Chassis configuration in the layer-2 test case. Layer-3 recovery times were somewhat higher due to the greater amount of route processing involved with OSPF, but again were well below the 1-second mark.
Figure 6: The Virtual Chassis Channel Capacity Validation Test Bed As shown in Figure 6, this configuration involved a total of 108 gigabit Ethernet test ports. The Spirent TestCenter generators offered a bidirectional aggregate load of up to 30 Gbit/s running between the lower two Virtual Chassis connections. This rate, 30-Gbit/s in each direction on each of two VCPs on the bottom switch in Figure 6, represents a practical upper bound for user data. Although the physical channel capacity is higher (around 32 Gbit/s per VCP in each direction), some bandwidth is used by frame headers added for internal switching.
Theoretical maximum Measured throughput Percentage of for 30 ports (fps) on 30 ports (fps) line rate 64 44,642,857 44,642,857 100.00% 128 25,337,838 25,337,838 100.00% 256 13,586,956 13,586,956 100.00% 512 7,048,872 7,048,872 100.00% 1,024 3,591,954 3,591,954 100.00% 1,280 2,884,615 2,884,615 100.00% 1,518 2,438,231 2,438,231 100.00%
10
Table 3: Virtual Chassis Channel Capacity Throughput The Virtual Chassis configuration forwarded all frame sizes at line rate, validating that Virtual Chassis connections were not a bottleneck in latency or resiliency testing.
Conclusion
These tests demonstrated several ways in which Virtual Chassis technology can enhance network performance. A Virtual Chassis configuration clearly reduces latency compared with a standalone configuration. Moreover, Virtual Chassis configurations provide sub-second recovery from link and equipment failures (and in some cases, recovery is well into the sub-millisecond range). Throughput testing also demonstrated the ability of Virtual Chassis links to carry user traffic at 30-Gbit/s rates with zero frame loss, regardless of frame length.
11
Appendix B: Disclaimer
Version 2010050400. Network Test Inc. has made every attempt to ensure that all test procedures were conducted with the utmost precision and accuracy, but acknowledges that errors do occur. Network Test Inc. shall not be held liable for damages which may result for the use of information contained in this document.
12
Network Test Inc. 31324 Via Colinas, Suite 113 Westlake Village, CA 91362-6761 USA +1-818-889-0011 http://networktest.com