# Interleaved Weighted Round-Robin: A Network Calculus Analysis

arXiv: https://arxiv.org/abs/2003.08372 · HTML: https://mlsanity.com/paper/2003.08372/

## Key facts

- Cited by 4 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 18 Mar 2020.

## Built on by (most cited first, 4 of 4)

- [SWARM Parallelism: Training Large Models Can Be Surprisingly Communication-Efficient](https://mlsanity.com/paper/2301.11913/) (2023)
- [Helix: Serving Large Language Models over Heterogeneous GPUs and Network via Max-Flow](https://mlsanity.com/paper/2406.01566/) (2024)
- [An Alternative Practice of Tropical Convolution to Traditional Convolutional Neural Networks](https://mlsanity.com/paper/2103.02096/) (2021)
- [Load Balancing for AI Training Workloads](https://mlsanity.com/paper/2507.21372/) (2025)

## Abstract (arXiv)

Weighted Round-Robin (WRR) is often used, due to its simplicity, for scheduling packets or tasks. With WRR, a number of packets equal to the weight allocated to a flow can be served consecutively, which leads to a bursty service. Interleaved Weighted Round-Robin (IWRR) is a variant that mitigates this effect. We are interested in finding bounds on worst-case delay obtained with IWRR. To this end, we use a network calculus approach and find a strict service curve for IWRR. The result is obtained using the pseudo-inverse of a function. We show that the strict service curve is the best obtainable one, and that delay bounds derived from it are tight (i.e., worst-case) for flows of packets of constant size. Furthermore, the IWRR strict service curve dominates the strict service curve for WRR that was previously published. We provide some numerical examples to illustrate the reduction in worst-case delays caused by IWRR compared to WRR.