[Paper Review] Forward Delay-based Packet Scheduling Algorithm for Multipath TCP

Tuan-Anh Le, Loc X. Bui|arXiv (Cornell University)|Jan 13, 2015

Network Traffic and Congestion Control12 references4 citations

TL;DR

This paper proposes the Forward Delay-based Packet Scheduling (FDPS) algorithm for Multipath TCP (MPTCP) to minimize out-of-order packet delivery by scheduling packets based on estimated forward delay differences between paths. Unlike prior methods relying on RTT or clock synchronization, FDPS estimates forward delay differences without requiring synchronized clocks, significantly reducing reorder buffer occupancy and outperforming FIFO, MTCS, and Kim-2012 in simulations across diverse network conditions.

ABSTRACT

Multipath TCP (MPTCP) is a transport layer protocol that allows network devices to transfer data over multiple concurrent paths, and hence, utilizes the network resources more effectively than does the traditional single-path TCP. However, as a reliable protocol, MPTCP still needs to deliver data packets (to the upper application) at the receiver in the same order they are transmitted at the sender. The out-of-order packet problem becomes more severe for MPTCP due to the heterogeneous nature of delay and bandwidth of each path. In this paper, we propose the forward-delay-based packet scheduling (FDPS) algorithm for MPTCP to address that problem. The main idea is that the sender dispatches packets via concurrent paths according to their estimated forward delay and throughput differences. Via simulations with various network conditions, the results show that our algorithm significantly maintains in-order arrival packets at the receiver compared with several previous algorithms.

Motivation & Objective

• To address the out-of-order packet problem in MPTCP caused by heterogeneous path delays and bandwidths.
• To design a packet scheduling algorithm that minimizes reordering buffer occupancy at the receiver without requiring clock synchronization.
• To improve MPTCP performance under dynamic and asymmetric network conditions, including background traffic and multiple paths.
• To develop a forward delay estimation technique that is robust to asymmetric path characteristics.
• To evaluate the proposed algorithm against existing schemes under various realistic network topologies and traffic loads.

Proposed method

• Proposes a novel technique to estimate forward delay differences between concurrent paths without requiring clock synchronization between sender and receiver.
• Uses one-way delay estimation based on RTT measurements and acknowledges asymmetric path behavior, avoiding assumptions of symmetric delays.
• Employs a scheduling strategy that prioritizes paths with lower estimated forward delay for sending packets to maintain in-order delivery.
• Integrates the delay estimation with congestion window availability to ensure efficient path utilization.
• Implements the FDPS algorithm in ns-2 and evaluates performance using the Reorder Buffer-occupancy Density (RBD) metric.
• Validates the method under multiple scenarios: two-path symmetric/asymmetric topologies, background TCP traffic, and three-path configurations.

Experimental results

Research questions

• RQ1 Can forward delay differences between MPTCP paths be estimated accurately without clock synchronization?
• RQ2 How does FDPS compare to FIFO, MTCS, and Kim-2012 in reducing reorder buffer occupancy under varying network conditions?
• RQ3 Does FDPS maintain superior performance when background traffic distorts RTT measurements?
• RQ4 How does the number of concurrent paths affect the reordering problem, and can FDPS scale effectively?
• RQ5 To what extent does FDPS reduce the risk of packet loss due to finite reorder buffers?

Key findings

• FDPS achieves the lowest mean reorder buffer occupancy across all experiments, with 2.2 ± 0.02 MSS in Exp. A1 and 24.5 ± 1.14 MSS in the three-path scenario.
• In Exp. A2 (highly asymmetric paths), FDPS reduces buffer occupancy to 2.2 ± 0.03 MSS, compared to 167.6 ± 0.06 MSS for FIFO and 22.7 ± 0.02 MSS for MTCS.
• Under background traffic (Exp. A4), FDPS maintains 8.4 ± 0.06 MSS buffer occupancy, outperforming Kim-2012 (16.5 ± 0.78 MSS) and MTCS (19.8 ± 0.42 MSS).
• In asymmetric topologies with ACK compression (Exp. A5), FDPS achieves 29.2 ± 2.40 MSS buffer occupancy, significantly lower than FIFO (1682 ± 57.1 MSS).
• FDPS consistently outperforms all baselines in terms of RBD, with results concentrated near zero, indicating high in-order packet delivery.
• The algorithm remains robust under high path diversity and dynamic traffic, demonstrating scalability and resilience to RTT measurement noise.