A Linked Slow-Start Algorithm for MPTCP

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, the subflows' slow-start phase remains unchanged to RFC 5681, and all the subflows at this stage behave like independent TCP flows. Following the development of IW as per , each MPTCP subflow can start 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 , 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 shares the congestion window MPTCP subflows in slow start whenever a new subflow joins.

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. Acronyms used in this document: Initial Window Round Trip Time Congestion Window MPTCP subflow's inflight data Congestion Window of the subflow having largest sending rate New incoming subflow's Congestion Window a boolean variable indicating whether ACKs should be ignored the number of ACKs for which old_subflow.CWND stops increasing during slow-start sum of CWND of the subflows in slow-start

Since 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. The MPTCP aggregate only diverges from this behavior when the number of subflows changes. Coupling of CWND is therefore not necessary in slow-start except when a new subflow joins.

We illustrate the problematic 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 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. This problem, and the improvement attained with LISA, are documented in detail in .

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 and 3 segments as the lower limit . This is based on , 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 . After 2 RTTs, the old_subflow.CWND = 40 and a new_subflow joins the connection. Since old_subflow.CWND >= 20 (refer to ), 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 (assuming a receiver without delayed ACKs). 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 a compound CWND of 80+20=100 after 1 RTT.

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 // take IW(10) packets old_subflow.CWND -= 10 new_subflow.CWND = 10 Ignore_ACKs = True else if old_subflow.CWND ≥ 6 // take half the packets new_subflow.CWND -= old_subflow.CWND / 2 old_subflow.CWND -= new_subflow.CWND Ignore_ACKs = True else new_subflow.CWND = 3 // can't take from old_subflow if Ignore_ACKs and Inflight > old_subflow.CWND // do not increase CWND when ACKs arrive ACKs_To_Ignore = Inflight - old_subflow.CWND

LISA is implemented as a patch to the Linux kernel 3.14.33+ and within MPTCP's v0.89.5. It is meant for research and provided by the Unviersity of Oslo and Simula Research Laboratory, and available for download from http://heim.ifi.uio.no/runabk/lisa This code was used to produce the test results that are reported in .

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.

This memo includes no request to IANA.

Changes made to this document: Some minor text improvements and updated a reference.