Virtual output queues

6300 series switches

The Aruba 6300 series switches contain 8 MB of packet buffer memory. Approximately 5 MB are used for buffering traffic arriving on network ports. Each port can buffer at most 1 MB of packets divided equally across all the active queues in a queue profile. For example, using the default eight-queue profile, each virtual output queue (VOQ) can buffer a maximum of 128 KB of packets.

Replicated traffic (L2 broadcasts, L2 multicasts, IP multicasts, mirrored packets, and unknown-unicasts) requires an extra step. There are eight additional replication queues initially used for replicated traffic. The switch services the replication queues strictly at the rate of 100 G. For ports on the same switch, pointers to the same packet are placed in the VOQ for each destination port. For ports on remote VSF member switches, only one copy of the packet data is moved across the VSF link to the destination replication VOQ to optimize link bandwidth. From there, the destination switch puts pointers to the same packet on the VOQ for each destination port.

6400 series switches

The 6400 series switches use a VOQ architecture where most packet buffering occurs on the ingress line module. Only shallow buffers are used on the egress line module. This optimizes usage of fabric bandwidth. Traffic destined for one port (unicast) uses different buffering and scheduling than traffic destined for multiple ports (replication).

Each line module contains eight VOQs for every destination port in the chassis, and eight egress physical queues (EPQs) for the ports local to the line module, regardless of the number of queues in the queue profile. Aruba 6400 Series R0X38-R0X45 line modules contain 8 MB of packet buffer memory. Approximately 5 MB is used for buffering traffic arriving on network ports. The remainder is used for buffering traffic arriving across the fabric. On each line module, at most 1 MB of buffering is available for every destination port in the chassis. This 1 MB buffer is equally divided across all the active queues in the queue profile.

For example, using the default eight-queue profile, each VOQ can buffer at most 128 KB of packets. Packets wait in VOQs on the arrival line module until the destination port scheduler selects them to be moved to one of its EPQs. If the destination port is on another line module, the packet is moved across the fabric. If the destination port EPQ is on the same line module, only a pointer to the packet is moved from the VOQ to the EPQ.

The schedule profile determines the order the destination port scheduler uses for VOQ servicing. The scheduler moves only enough packets to EPQs as are needed to keep the destination port transmitting at the configured rate.

Replicated traffic (L2 broadcasts, L2 multicasts, IP multicasts, mirrored packets, and unknown-unicasts) requires an extra step. There are eight additional internal replication VOQs per line module initially used for replicated traffic. The line module services these replication VOQs strictly at a rate of 100 G. For ports on the same line module, pointers to the same packet are placed in the VOQ for each destination port. For ports on remote line modules, only one copy of the packet data is moved across the fabric to the destination replication VOQ, to optimize fabric bandwidth. From there, the destination line module will put pointers to the same packet on the VOQ for each destination port.