A super-aggregation strategy for multi-homed mobile hosts with heterogeneous wireless interfaces

Abstract

Most mobile devices today are equipped with multiple and heterogeneous wireless interfaces. In this paper we ask the following question: What is the best approach to leverage the multiple interfaces available at a mobile device in terms of the performance delivered to the user? In answering the question we argue that simple “bandwidth aggregation” approaches do not provide any meaningful benefits when the multiple interfaces used have highly disparate bandwidths as is true in many practical environments. We then present super-aggregation, a set of mechanisms that in tandem use the multiple interfaces intelligently and in the process is able to achieve a performance that is “better than the sum of throughputs” achievable through each of the interfaces individually. We prototype super-aggregation on both a laptop and the Google Android mobile phone and demonstrate the significant (up to 3$\times$ throughput) performance improvements it provides in real-world experiments. We conduct both theoretical analysis and extensive experiments to show that super-aggregation is able to improve throughput beyond the sum of the parts under most of the cases.

Appendices

Appendix: Super-aggregation beyond TCP

1.1 Generic principles and case study

We now generalize the super-aggregation principles proposed in Sect. 3 and demonstrate their generic application with a case study of a non-TCP protocol. Each generic principle describes an approach of leveraging multiple interfaces of wireless devices that can increase throughput beyond the sum of the parts. We pick rate-adaptive video streaming [41] for the case study since it is a popular UDP application in the Internet. In such systems, the server sends video streams in the form of UDP datagrams to clients. To provide good video quality in response to capacity variation, the server adjusts its codec or sending rate based on available bandwidth to the client.

1.1.1 Selective offloading: tackling self-contention of client reports

Selective offloading chooses some packets to move from the WiFi interface to the 3G interface. One design principle is moving some packets to the 3G interface to resolve self-contention in the WiFi network, as in offloading-ACK. Moving small packets can provide significant improvements since overhead of sending them via the WiFi interface is relatively higher. The other design principle is to use the 3G interface when overall performance is affected by some characteristics that the WiFi interface performs poorer than the 3G interface. In rate-adaptive video streaming, clients keep sending reports of traffic characteristics to the sender to rate adaptation. The rate adaptation and overall throughput may be impaired by intermittent availability of the WiFi interface of a mobile host⁵ On the other hand, 3G provides better availability in terms of broader coverage of each base station and thus fewer disconnections to the same mobile host. A video client on mobile host is unable to send reports via a WiFi link under blackout and handoff, and the server may interpret missing of reports as client disconnection or network congestion. Incorrect characterization causes improper rate adaptation of the video stream. Offloading-report moves report packets to the 3G interface with high availability enables continuous reporting, which allows server to do timely and accurate rate adaptation. It also reduces packet loss rate of video on downlink, since small report packets cause self-contention in WiFi networks.

1.1.2 Proxying: improving reliability and timeliness of command packets

Proxying improves performance of the connection on WiFi by masquerading packets from 3G, which serves as a proxy when WiFi is temporarily unavailable. One design principle is to enable communication when the WiFi interface has blackouts, as in proxying-blackout-freeze. Adding control packets via the 3G interface can help in preventing blackout’s adverse effects to application. The other design principle is to improve reliability and timeliness of some packets by sending them to both interfaces. Heterogeneous interfaces provide diversity in packet losses, so sending a redundant packet to the 3G interface can effectively improve end-to-end reliability. Video streaming clients send command packets to perform control operations, such as pause/resume video delivery and updating configurations. Losing command packets degrades response time perceived by the client. Proxying-redundant-commands sends a duplicate copy of those command packets to improve the reliability. It improves response time to the client and may also improve other dimensions of performance since commands are delivered more timely.

1.1.3 Mirroring: reducing loss rate of baseline frames

Mirroring creates a independent connection on the 3G interface, and same or related content is downloaded to improve performance by leveraging loss diversity. One design principle is decoupling some of high-layer mechanisms to the mirroring connections, as in mirroring-loss-fetching. This helps the client to separate operation of two mechanisms that have adverse interaction in WiFi networks. The other design principle is reducing packet loss rate by fetching redundant contents from both connections, especially essential portions in the original connection. Scalable Video Coding [42] is commonly considered in rate-adaptive video streaming since it encodes video content into different quality levels in a scalable way. All clients receive baseline frames and those with higher capacity also receive enhancement frames that rely on baseline frames. Baseline frames are more critical packets and requires less bandwidth than enhancement frames. Mirroring-baseline-frames establishes a mirroring connection via the 3G interface and requests for a baseline video stream of the same video content. Having duplicate baseline frames from server can significantly improve overall video quality, especially in a lossy environment.

1.2 Generic architecture

We present the generalized software architecture in a modular fashion such that it is evident how to reuse the common components for generic super-aggregation principles. As shown in Fig. 3(a), some components in the super-aggregation architecture are specific for TCP, while others provide common functionalities needed in other principles. For example, Offloader is used for offloading-ACK, Report Offloading, and Voice Offloading. Offloading-ACK uses ACK marker to notify Offloader that ACK packets should be offloaded. The Report Offloading and Voice Offloading should have their own component to mark report packets and voice frames. The Offloader takes all packets marked and split them according to available uplink bandwidth on 3G interface. Other common components include the Blackout Detector and the Interface Characterizer Table (Table 2).

Table 2 Variables

Theoretical analysis

In this section we present an analytical model to study the throughput of super-aggregation. The analysis answers a fundamental question: in what conditions does super-aggregation yield throughput more than sum of the parts? The analytical model can also be used to assess the performance of super-aggregation when applied to other networks. Each of the three super-aggregation principles are analyzed separately in the corresponding network conditions. In the analysis of offloading-ACK, there is no blackout or random wireless loss. In the analysis of proxying-blackout-freeze, there are blackouts but no random wireless loss. In the analysis of mirroring-loss-fetching, there are random wireless losses but no blackout. We summarize the analysis of the average throughput of default TCP and each super-aggregation principle. The detailed analysis is included in [8] due to lack of space. At the end of this section, we presents the insights from the analysis to show that in most of the cases super-aggregation is able to provide throughput more than sum of the parts. The analytical model is validated with experiments in the next section.

1.1 Analysis of offloading-ACK

When self-contention occurs to TCP, uplink ACK packets compete with downlink data packets for the same wireless network resources. A TCP receiver replies with an ACK for every two data segments. The saturated throughput of TCP at the wireless link is determined by the time to send both the TCP data segments and ACKs (Table 3).

Table 3 Parameters

$$\begin{aligned} T_s(tcp) = \frac{2mss}{2t_{data} + t_{ack}} \end{aligned}$$

(1)

Offloading-ACK eliminates self-contention of TCP in the WiFi network by moving TCP ACKs to the 3G link. With the WiFi link purely used for downstream TCP data segments, the saturated throughput does not contain time to send TCP ACKs.

$$\begin{aligned} T_s(offloading) = \frac{mss}{t_{data}} \end{aligned}$$

(2)

We provide a more detailed analysis of the average throughput of TCP based on its congestion control mechanism.

$$\begin{aligned} T(p) \cong \frac{3}{4}\left( \frac{T_s \times RTT}{mss}+r_p\right) \times \frac{mss}{RTT} \end{aligned}$$

(3)

1.2 Analysis of proxying-blackout-freeze

To analyze the impact from blackout, we assume that blackouts occurred in the WiFi network periodically. The period of a blackout is $t_{blackout}$, and the frequency of its occurrence is $f_{blackout}$.

$$\begin{aligned} T(tcp) = \frac{t_{ss} \times \frac{(W_s + 1)mss}{2RTT} + (t_{interval} - t_{idle} - t_{ss}) \times \frac{(W_{max} + W_{min})mss}{2RTT}}{t_{interval}} \end{aligned}$$

(4)

Proxying-blackout-freezing can freeze the TCP transmission during the blackout and immediately resumes the transmission after the blackout.

$$\begin{aligned} T(proxying) = \frac{(t_{interval} - t_{blackout}) \times \frac{(W_{max} + W_{min})mss}{2RTT}}{t_{interval}} \end{aligned}$$

(5)

1.3 Analysis of mirroring-loss-fetching

To analyze the impacts from random wireless loss, we assume that each packet loss is an independent event with probability $p_l$. The analysis of TCP throughput is based on $N$, the number of packets sent in a congestion avoidance phase, which is defined as the duration between two packet losses.

$$\begin{aligned}&W_{max}(p_l)=\sqrt{\frac{8}{3}E[N(p_l)]+3}-1 \end{aligned}$$

(6)

$$\begin{aligned}&T(tcp, p_l) \cong \frac{3}{4}W_{max}(p_l)\times \frac{mss}{RTT} \end{aligned}$$

(7)

Mirroring-loss-fetching prevents TCP from unnecessary reduction in congestion window by hiding random wireless losses to the TCP sender. The throughput achieved with mirroring-loss-fetching is constrained by two factors. The first one is the throughput of data transfer in the primary connection, and the second one is the rate in recovering lost packets in the mirroring connection. The rate in recovering lost packets $T_r$ is determined by the average time to trigger a data segment (no matter lost or not) from the TCP sender.

$$\begin{aligned} T_r&= \frac{mss}{p_l \times t_{g,data} + (1-p_l) \times t_{g,ack}} \end{aligned}$$

(8)

$$\begin{aligned} T(mirroring, p_l)&= min\big (T(tcp, 0), T_r\big ) \end{aligned}$$

(9)

1.4 Insights from the analysis

In this subsection, we present the insights from the theoretical analysis on the throughput of super-aggregation under different conditions. We specifically look at an important aspect of the degree of heterogeneity required in the capacity of the two interfaces to observe aggregate throughput more than sum of the parts. The following observations will be verified with extensive experiments in the next section.

The throughput improvement from offloading-ACK and/or proxying-blackout-freeze is independent from the capacity of the secondary interface. As shown in the theoretical analysis, the throughput formulas of offloading-ACK (Eq. 3) and proxying-blackout-freeze (Eq. 5) do not contain the capacity of the secondary interface. Those two principles are able to achieve a saturated throughput higher than sum of the parts no matter how small the capacity of the secondary interface is. While the secondary interface has higher capacity than necessary, offloading-ACK or proxying-blackout-freeze does not exhaust its capacity, and the remaining capacity can be utilized with simple aggregation. In general, offloading-ACK and proxying-blackout-freeze can provide aggregate throughput more than the sum of the parts under any degree of capacity heterogeneity.

Contrary to the above two principles, the throughput improvement of mirroring-loss-fetching depends on the capacity of the secondary interface. Eq. 9, when capacity of the secondary interface is significantly lower than that of the primary interface, the overall throughput of mirroring-loss-fetching is dominated by $T_r$. In that case, the aggregate throughput depends on the capacity of the secondary interface, as shown in Eq. 8. That is because the data transport of mirroring-loss-fetching is stagnated by waiting for loss recovery in the secondary interface. The analytical model can also be used to predict if mirroring-loss-fetching is beneficial when the capacity in the secondary interface drops down. If the capacity in the secondary interface is insufficient to benefit the overall throughput, super-aggregation should not hide all packet loss in the primary connection.

1.5 Validation of the analysis

In this subsection, we conduct extensive experiments with diverse capacity in the two interfaces to validate the theoretical analysis. To manipulate the network capacity in the experiment testbed, we use a WiFi AP (Linksys WAP55AG) and an emulated cellular network (Network Nightmare WAN Emulator). By using the network emulation, we covers a broad range of wireless technologies of heterogeneity capacity. The data rate in the WiFi interface ranges from 6Mbps to 54Mbps. The data rate of the secondary interface also varies from 2.5G (GPRS/EDGE) with around 100kbps to 4G (WiMAX/LTE) with tens of Mbps.

Figure 12 compares the analysis and experimental results of both default TCP and offloading-ACK. The results show that the analytical model accurately capture the throughput performance of TCP and offloading-ACK under different circumstances. It is noteworthy in Fig. 12(b) that when the cellular network has capacity lower than 500kbps, the observed throughput in our default experiment settings is lower than what expected by the analysis. The reason is because offloading ACKs through a cellular interface of narrow bandwidth slows down TCP in achieving the saturated throughput, even though the saturated throughput is independent of the cellular interface capacity. RFC 3465 [43] specifies that TCP should not increase its congestion window by more than two full segments with each acknowledgement, the growth rate in congestion control window is reduced when fewer acknowledgements are delivered. With the 100-s data transfer in default experiments, TCP with offloading-ACK does not have enough time to achieve the supported throughput. By extending the connection duration to 1000 s, the average throughput achieved in the experiments matches with the analysis. From the analytical framework, we have derived the maximum data usage of offloading-ACK in 3G as 317Kbps. It matches with the above experimental observation. Such data usage is considered low when compared with the throughput improvement of 11.27 Mbps in the WiFi network.

Fig. 12

Comparison of offloading-ACK analysis and experiments

Figure 13 compares the analysis and experimental results of both default TCP and proxying-blackout-freeze under blackouts. The default blackout duration is 2 s. The analysis and the experimental results are very close to each other. The analysis of default TCP has higher throughput than experiments. That can be caused by the overhead in TCP retransmission in practice, such as a coarse-grained timer for TCP timeout.

Fig. 13

Comparison of proxying-blackout-freeze analysis and experiments

Figure 14 compares the analysis and experimental results of both default TCP and mirroring-loss-fetching under random wireless loss. The default packet loss rate is 0.1 %. The matching of the analysis and the experimental results shows that the analytical model is very accurate in estimating the throughput under random wireless loss.

Fig. 14

Comparison of mirroring-loss-fetching analysis and experiments