Infiniband switches support multiple virtual channels per port which allow traffic types to be treated differently based on class of service indicators. Unlike Ethernet, no routing or address learning features are required, which simplifies the design and allows Infiniband switches to achieve very low cut-through latencies on the order of 100-200 nS. Forwarding tables simply consist of destination address-output port pairs that are populated during initialization and may be modified if something changes in the fabric. Infiniband switches have less functional overhead than Ethernet switches in that they simply forward packets based on header address information. Although this eliminates the possibility of checking the frame for errors using the frame check sequence before transmission, most fabrics of this type expect that the fabric interface adapters on the receiving compute nodes will perform this check when the packet is received and flag any errors. Transmission of the frame can then start before the entire frame is received. In this mode of operation, incoming frame headers are inspected and forwarding decisions can be made in a few hundred nanoseconds. Gary Lee, in Cloud Networking, 2014 SwitchĮthernet switches that are used in HPC applications use cut-through operation in order to reduce latency.
0 kommentar(er)
