Multipath TCP R. Barik
Internet-Draft University of Oslo
Intended status: Standards Track S. Ferlin
Expires: April 21, 2016 Simula Research Laboratory
M. Welzl
University of Oslo
October 19, 2015

A Linked Slow-Start Algorithm for MPTCP


This document describes the LISA (Linked Slow-Start Algorithm) for Multipath TCP (MPTCP). Currently during slow-start, subflows behave like independent TCP flows making MPTCP behave unfairly to cross-traffic and causing more congestion in the bottleneck, which yields more losses among the MPTCP subflows. LISA couples the initial windows (IW) of MPTCP subflows during the initial slow-start phase to remove this adverse behavior.

Status of This Memo

This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.

Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at

Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."

This Internet-Draft will expire on April 21, 2016.

Copyright Notice

Copyright (c) 2015 IETF Trust and the persons identified as the document authors. All rights reserved.

This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents ( in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.

Table of Contents

1. Introduction

MPTCP is an ongoing standardization effort that aims to extend TCP by allowing multiple paths to be used simultaneously. The current MPTCP implementation provides multiple congestion control algorithms, which aim to provide fairness to TCP flows at the shared bottlenecks. However, in RFC 6356 [RFC6356], the subflows' slow-start phase remains unchanged to RFC 5681 [RFC5681], and all the subflows at this stage behave like independent TCP flows. Following the development of IW as per [RFC6928], each MPTCP subflow starts with IW = 10. With an increasing number of subflows, the subflows' collective behavior during the initial slow-start phase can temporarily be very aggressive towards a concurrent regular TCP flow at the shared bottleneck.

According to [UIT02], most of the TCP sessions in the Internet consist of short flows, e.g., HTTP requests, where TCP will likely never leave slow-start. Therefore, the slow-start behavior becomes of critical importance for the overall performance.

To mitigate the adverse effect during initial slow-start, we introduce LISA, the "Linked Slow-Start Algorithm". LISA's design is based on initial congestion window sharing of MPTCP subflows, hence, providing coupling in the window increase.

1.1. Definitions

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119 [RFC2119].

Acronyms used in this document:

IW --
Initial Window
RTT --
Round Trip Time
Congestion Window
Inflight --
MPTCP subflow's inflight data
old_subflow.CWND --
Congestion Window of the subflow having largest sending rate
new_subflow.CWND --
New incoming subflow's Congestion Window
Ignore_ACKs --
a boolean variable indicating whether ACKs should be ignored
ACKs_To_Ignore --
the number of ACKs for which old_subflow.CWND stops increasing during slow-start
compound CWND --
sum of CWND of the subflows in slow-start

2. MPTCP Slow-Start Problem Description

Given that it takes 1 RTT for the sender to receive any feedback on a given TCP connection, sending an additional segment after every ACK is rather aggressive. Therefore in slow-start, all subflows independently doubling their CWND as in regular TCP, results in MPTCP also doubling its compound CWND.

2.1. Example of current MPTCP slow-start problem

We illustrate the MPTCP slow-start behavior with an example: Consider an MPTCP connection consisting of 2 subflows. The first subflow starts with IW = 10, and after 2 RTTs the CWND becomes 40 and a new subflow joins, again with IW = 10. Then, the compound CWND becomes 40+10 = 50. With an increasing number of subflows, the compound CWND in MPTCP becomes larger than that of a concurrent TCP flow.

For example, MPTCP with eight subflows (as recommended in [DCMPTCP11] for datacenters) will have a compound CWND of 110 (40+7*10). As a result, MPTCP would behave unfairly to a concurrent TCP flow sharing the bottleneck. This aggressive behavior of MPTCP also affects the performance of MPTCP. If multiple subflows share a bottleneck, each of them doubling their rate every RTT, will cause excessive losses at the bottleneck. This makes MPTCP enter the congestion avoidance phase earlier and thereby increases the completion time of the transfer.

3. Linked Slow-Start Algorithm

3.1. Description of LISA

The idea behind LISA is that each new subflow takes a 'packet credit' from an existing subflow in slow-start for its own IW. We design the mechanism such that a new subflow has 10 segments as the upper limit [RFC6928] and 3 segments as the lower limit [RFC3390]. This is based on [RFC6928], [RFC3390] and the main reason behind it is to let these subflows compete reasonably with other flows. We also divide the CWND fairly in order to give all subflows an equal chance when competing with each other.

LISA first finds the subflow with the largest sending rate measured over the last RTT. Depending on the subflow's CWND, between 3 and 10 segments are taken from it as packet credit and used for the new subflow's IW. The packet credit is realized by reducing the CWND from the old subflow and halting its increase for ACKs_To_Ignore number of ACKs.

We clarify LISA with the example given in Section 2.1. After 2 RTTs, the old_subflow.CWND = 40 and a new_subflow joins the connection. Since old_subflow.CWND >= 20 (refer to Section 3.2), 10 packets can be taken by the new_subflow.CWND, resulting in old_subflow.CWND = 30 and new_subflow.CWND = 10. Hence, MPTCP's compound CWND, whose current size is 40, should ideally become 60+20 = 80 after 1 RTT. (Linux sends ACKs for every segment in slow-start.) However, if 40 segments from old_subflow.CWND are already in flight, the compound CWND becomes in fact 70+20 = 90. Here, LISA keeps old_subflow.CWND from increasing for the next 10 ACKs. In comparison, MPTCP without LISA would have 80+20=100 after 1 RTT.

3.2. Algorithm

Below, we describe the LISA algorithm. LISA is invoked before a new subflow sends its IW.

Before computing the new_subflow.CWND, Ignore_ACKs = False and ACKs_To_Ignore = 0.
Then, ignoring the new_subflow, the subflow in slow-start with the largest sending rate (old_subflow.CWND, measured over the last RTT) is selected.
If there is no such subflow, the IW of the new_subflow.CWND = 10 Otherwise, the following steps are executed:
if old_subflow.CWND ≥ 20
old_subflow.CWND -= 10
new_subflow.CWND = 10
Ignore_ACKs = True
else if old_subflow.CWND ≥ 6
new_subflow.CWND -= old_subflow.CWND / 2
old_subflow.CWND -= new_subflow.CWND
Ignore_ACKs = True
new_subflow.CWND = 3
if Ignore_ACKs and Inflight > old_subflow.CWND
// do not increase CWND when ACKs arrive
ACKs_To_Ignore = Inflight - old_subflow.CWND

4. Implementation Considerations

LISA is implemented as a patch to the Linux kernel 3.14.33+ and within MPTCP's v0.89.5.

5. Conclusions

We identify the adverse effect of MPTCP's uncoupled slow-start on the performance of MPTCP itself and on concurrent TCP traffic. We propose LISA, a linked slow-start algorithm for MPTCP that couples MPTCP subflows during slow-start phase. LISA was implemented as a patch to the Linux kernel and evaluated in both emulated and real testbeds [lisa]. In this evaluation, we observed that TCP (CUBIC) completes its transmission earlier than MPTCP without LISA. This is due to the large overshoot when an additional subflow joins, causing more retransmissions. LISA solves this problem.

6. Acknowledgements

This work was part-funded by the European Community under its Seventh Framework Programme through the Reducing Internet Transport Latency (RITE) project (ICT-317700). The authors also would like to thank David Hayes (UiO) for his comments. The views expressed are solely those of the authors.

7. IANA Considerations

This memo includes no request to IANA.

8. Security Considerations

9. Change History

Changes made to this document:

00->00 :
First version

10. References

10.1. Normative References

[RFC2119] Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, March 1997.
[RFC3390] Allman, M., Floyd, S. and C. Partridge, "Increasing TCP's Initial Window", RFC 3390, DOI 10.17487/RFC3390, October 2002.
[RFC5681] Allman, M., Paxson, V. and E. Blanton, "TCP Congestion Control", RFC 5681, DOI 10.17487/RFC5681, September 2009.
[RFC6356] Raiciu, C., Handley, M. and D. Wischik, "Coupled Congestion Control for Multipath Transport Protocols", RFC 6356, DOI 10.17487/RFC6356, October 2011.
[RFC6928] Chu, J., Dukkipati, N., Cheng, Y. and M. Mathis, "Increasing TCP's Initial Window", RFC 6928, DOI 10.17487/RFC6928, April 2013.

10.2. Informative References

[DCMPTCP11] Raiciu, C., Barre, S., Pluntke, C., Greenhalgh, A., Wischik, D. and M. Handley, "Improving datacenter performance and robustness with multipath TCP", ACM SIGCOMM p266-277, August 2011.
[lisa] Barik, R., Welzl, M., Ferlin, S. and O. Alay, "LISA: A Linked Slow-Start Algorithm for MPTCP", the paper will be available as soon as possible , 2015.
[UIT02] Brownlee, N. and K. Claffy, "Understanding internet traffic streams: Dragonflies and tortoises", IEEE Communications Magazine p110-117, 2002.

Authors' Addresses

Runa Barik University of Oslo PO Box 1080 Blindern Oslo, N-0316 Norway EMail:
Simone Ferlin Simula Research Laboratory P.O.Box 134 Lysaker, 1325 Norway EMail:
Michael Welzl University of Oslo PO Box 1080 Blindern Oslo, N-0316 Norway Phone: +47 2285 2420 EMail: