Performance evaluation of open virtual routers

Data Plane Optimization in Open Virtual Routers

“…However, for the purpose of redirecting packets between virtual and physical interfaces, it introduces a considerable amount of overhead. This, in turn, incurs performance penalties, something that was investigated in previous work [9]. Similar solutions that have been used are the virtual switch [14] and the short bridge [5].…”

“…This functionality can be implemented in several ways. One common configuration is to use the regular software bridge module in the Linux kernel to interconnect the interfaces, as shown in [9]. The software bridge provides a general-purpose switching function that allows packets to be switched between interfaces (virtual or physical) based on MAC addresses and MAC address learning.…”

“…In previous work we investigated how a careful selection of virtual interface and approach for redirecting packets between virtual and physical interfaces can improve the overall system performance [9]. We compared a macvlan-based virtual router with veth/bridge-based virtual router using both OpenVZ and Linux namespaces virtualization environments.…”

“…Backlog queue congestion can cause serious throughput degradation for bridge/veth based virtual routers [9]. The backlog queue is maintained on a per CPU basis, which means that virtual routers running on the same CPU will share the same queue.…”

“…In previous work we have shown that this is a costly solution in terms of performance, as it introduces considerable processing overhead [9]. A more promising approach is to use the macvlan device -a kernel object specifically designed for support of multiplexing and demultiplexing of packets between physical and virtual interfaces based on MAC addresses.…”