Comment
Author: Admin | 2025-04-28
Switches come with intelligent buffer capabilities such as approximate fair drop (AFD). You can use AFD to distinguish high bandwidth (elephant flows) from short-lived and low bandwidth flows (mice flows). After AFD has information about which traffic makes up the elephant flows, AFD can mark ECN bits with 0x11 values, but only for high bandwidth flows. Based on the bandwidth used by the flow, a different number of packets is marked with ECN. For example, a flow running at 1G has fewer packets marked with 0x11 ECN bits than a flow running at 10G, so AFD triggers a different number of marked packets proportional to the size of the flow. With the right algorithms on the end hosts, using marking in this way can create an efficient way to slow the flows that contribute the most to creating congestion in the overall system. This way, performance is optimized for the lowest latency. An advantage AFD has over WRED is its ability to distinguish which set of flows are causing the most congestion. WRED marks all traffic in a queue equally. AFD is more granular and marks only the higher bandwidth elephant flows while leaving the mice flows unmarked so as not to penalize them or cause them to slow. In an AI cluster, it is advantageous to let short-lived communications run to completion by not allowing a long transfer of data and any resulting congestion to slow them down. Packet drops are still avoided, but many transactions get completed faster
Add Comment