Keysight: flowlet balancing evens out AI fabric traffic where ECMP piles it up

Keysight published a short lab study on load balancing in AI data-center fabrics (Eric Yu, Keysight blog, 27 May 2026). Short version: the traffic of a training job breaks classic ECMP, and flowlet-based balancing repairs it without the price of per-packet spraying.

Why ECMP suffers

All-Reduce and similar collectives fire bursts from every node at the same moment. ECMP hashes a whole flow onto one path and keeps it there. When 8 flows burst in sync, 2 of them land on the same link, that link queues, and the neighbouring links sit idle. Per-packet spraying fills every link but reorders packets, so the NIC has to reassemble and the transport may retransmit.

What flowlet does

A flowlet is one burst inside a flow. Between bursts there is a gap. If the gap is longer than the configured timeout, the switch may send the next burst down another equal-cost path. Packets inside a burst stay together, so nothing gets reordered.

Numbers from the test

8-way ECMP fabric, full link rate, 256 µs bursts:

• ideal share: 12.5% per path
• ECMP with synchronised bursts: up to 25% on some paths, close to zero on others
• flowlet balancing on the same traffic: near 12.5% on every path
• continuous, non-bursty traffic: flowlet gains are smaller and less stable

Keysight calls this a baseline: symmetric topology, no failed links, no incast, no interaction with PFC or DCQCN. Those are the next experiments.

Why it matters to our customers

Switch ASICs and NICs already ship with flowlet modes. Only a test bed shows whether a given box uses them well: the result depends on the timeout value, burst length and what happens to the fabric when a link goes down. Tests of this kind are what our cloud lab is for. The traffic generator is picked for the setup, Keysight, Spirent or Xena, and agreed with you before the test. Test. Compare. Decide.

Source: Flowlet Load Balancing for AI Workloads, Keysight Technologies.