TCP slow start with fair share of bandwidth

An accurate model for analyzing wireless TCP performance with the coexistence of Non-TCP traffic

BBR, a congestion control algorithm proposed by Google, regulates the source sending rate by deriving an estimate of the bottleneck’s available bandwidth and RTT of the path. The initial version of BBR, called BBRvl, was found to be unfair, getting higher than the fair share of bandwidth when co-existing on bottleneck links with other congestion control algorithms. It also does not perform as well with networks having routers with shallow buffers. To overcome these concerns, a newer version, called BBRv2, has been proposed. Our goal in this paper is to understand the differences between the two versions and examine the primary reasons behind the improvement in performance of BBRv2. We present an experimental evaluation of BBRvl and BBRv2, evaluating their fairness across connections using the same protocol (intra-protocol fairness) and using different protocols (inter-protocol fairness) as well as delay and link utilization. From experiments with shallow and deep buffers, BBRv2 is most effective when it uses Explicit Congestion Notification (ECN), but fairness issues continue to exist in BBRv2 when ECN is disabled. A concern for BBRv2 is that it is somewhat complex to deploy in Wide Area Networks (WAN) because of the dependency with the DCTCP-style reduction of the congestion window, which is primarily usable in low-feedback delay Data Center Networks.

In mobile ad hoc networks, one might expect multipath routing to provide some robustness to link failures and facilitate the transmission of packets along paths that avoid regions of congestion. Consequently, one would expect an improvement in network performance in terms of the achieved throughput. We consider TCP goodput as the metric of performance. We find that, contrary to expectations, not all TCP connections enjoy the benefits of multipath routing. Specifically, we find that while long (in terms of hop-count) TCP connections seem to benefit, short connections in fact suffer a slight degradation in goodput as compared to TCP using the single shortest path. Furthermore, we find that alternate path routing, wherein packets are routed on a secondary alternate path only upon the failure of the primary path, helps achieve almost the same goodput as when the multiple paths are used simultaneously. The main benefits of currently proposed multipath routing schemes seem to be limited to improving the efficiency of route discoveries that are initiated either due to real route failures (due to mobility) or due to false failures (due to interference effects) for long TCP connections.

Standard TCP reduce its performance by misinterpreting mobility losses due to node motion as congestion losses in ad hoc network. By far, there were some approaches proposed to improve TCP performance in ad hoc network. In general, they freeze TCP connection temporarily when route is failed, and resume TCP transmission from the original frozen state when route is reconstructed. In this paper, we propose a new approach, TCP-RC, which re-compute cwnd and ssthresh for the TCP connection after route is reconstructed, thus it can adjust the TCP transmission rate adaptively according to current capacity of TCP connection. Consequently, TCP-RC lowers possibility of bursty traffic and avoids invoking congestion control under situation of high network load. Analyses and simulations show that TCP-RC had achieved better TCP performance than Reno in ad hoc network.

This paper describes a congestion control method for TCP that adjusts the transmission rate of a TCP connection not only by changing the congestion window size as in normal TCP, but also by delaying the transmission of packets at the sender. We refer to this mechanism as TCP with sender-based delay control, or simply SDC. SDC can keep the window size of a TCP connection above a certain threshold even when its fair share of bandwidth is arbitrarily small. Since TCP fast retransmit and recovery is likely to work when the window size of the connection is sufficiently large, the new scheme can result in reduced-frequency of TCP timeouts for the connection. In particular, SDC allows many TCP flows to share a link without experiencing many timeouts. In addition, SDC reduces a well-known TCP bias against connections with large round trip times. The paper presents the principle behind SDC and simulation results demonstrating its properties and advantages.

With the rise of heterogeneous congestion control algorithms and increasingly complex application control loops (e.g. adaptive bitrate algorithms), the Internet community has expressed growing concern that network bandwidth allocations are unfairly skewed, and that some Internet services are 'winners' at the expense of 'losing' services when competing over shared bottlenecks. In this paper, we provide the first study of fairness between live, end-to-end services with distinct workloads. Rather than focusing on individual components of an application stack (e.g., studying the fairness of an individual congestion control algorithm), we want to provide a direct study over real-world deployed applications. Among our findings, we observe that services typically achieve less-than-fair outcomes: on average, the 'losing' service achieves only 72% of its max-min fair share of link bandwidth. We also find that some services are significantly more contentious than others: for example, one popular file distribution service causes competing applications to obtain as low as 16% of their max-min fair share of bandwidth when competing in a moderately-constrained setting.

TCP performs slow start phase based on an exponential increase of the congestion window to predict the current network. Current TCP slow start does not work well under large bandwidth networks. The exponential growth of slow start phase cause burst packet loss if there are many traffic in the network. In this paper, we use a statistical slow start method to solve heavy packet loss from collision between flows in the network and to improve TCP performance.

This paper proposes a partial state Active Queue Management (AQM) technique, called CHOKe-FS, which addresses the problem of fair-bandwidth allocation among different flows sharing a common bottleneck link. CHOKe-FS is a packet-dropping mechanism for classifying and restricting unresponsive or misbehaving flows during congestion. It utilizes a direct bitmap technique, which estimates the number of active flows in the buffer to compute the fair-share. It inherits the matched drop mechanism of CHOKe and the buffer management technique of RED. The proposed technique differs from CHOKe in three ways: (i) it divides the queue into multiple regions, (ii) it dynamically adjusts the drawing factor depending on the average queue occupancy, by drawing multiple packets from the buffer, and (iii) it estimates the number of active flows in a queue for fair share estimation to drop the packet from flows exceeding the fair bandwidth share. The performance of CHOKe-FS is analyzed under various scenarios by comparing it with the well-known, AQM schemes such as RED, CHOKe and CHOKeR. The results demonstrate that the proposed algorithm not only protects the responsive flows from unresponsive flows, but it also works reasonably well for providing intra-protocol fairness.

This paper analyzes a TCP connection over UMTS radio access bearer in RLC acknowledged mode. The stacking of RLC and TCP protocols is highly beneficial for the global system behavior. Protocol mechanisms used by RLC are well adapted to wireless constraints and the ARQ function implemented by RLC hides the wireless channel unreliability to TCP. However, RLC protocol parameters have to be carefully tuned otherwise the TCP performance can degrade due to RLC. One of the factors that can influence the TCP performance is the RLC buffer size. We analyzed some of factors that affect the RLC buffer occupancy and we show that an insufficient buffer size degrades the TCP performance.

In these days, IP network customers become sensitive to the QoS guarantee provided by the network. Amongst many QoS guarantee schemes, Diffserv (Differentiated Services) is the one of the most practical candidates for the next generation IP networks. Under these circumstances, the commercial network venders have already provided the Diffserv routers that support EF (Expedited Forwarding) PHB (Per Hop Behavior). However, AF (Assured Forwarding) PHB in Diffserv still has not been provided by commercial routers. Therefore, it is a realistic solution that SBR3 (Statistical Bit Rate 3) of ATM emulates AF PHB, but it is not clear whether TCP traffic over AF PHB emulated by ATM is differentiated from the best effort TCP traffic over DF (Default Forwarding) PHB. To confirm the differentiation, we have experimentally studied TCP performance through the link into which TCP connections over AF PHB and DF PHB are aggregated. This paper describes the experimental results and discusses the possibility of the TCP performance differentiation between AF PHB and DF PHB over ATM.

We show that different flavors of TCP may be viewed as implementations of age-based scheduling disciplines. By parameterizing the scheduling disciplines of interest we are able to position variants of TCP in a wide spectrum having FCFS (first-come first-served) and LAS (least attained service first) as extremal policies, and including PS (processor sharing) as an intermediate case. We argue that for highly loaded systems, providing a fair bandwidth allocation among all users is secondary to ensuring network stability. So as to isolate protocol fairness from congestion effects, we therefore focus on scenarios with infinite buffers. This way, asymmetries in capacity shares are the consequences of the protocol only, and not affected by the packet loss process. The model, however, is flexible enough to include finite buffers with random packet loss as a special case (for example to capture Active Queue Management). The results are helpful in studying fairness and performance concerned with transmission protocols in communication networks. For persistent HTTP connections we study the distributions of the transmission rates and the relative fairness index under various assumptions on the file size distributions and scheduling disciplines. For infinite file sizes, we show that protocols that increase priority more than linearly with the attained service asymptotically behave similar to FCFS. In contrast, protocols with at most linearly increasing (or even decreasing) priority, with TCP's Congestion Avoidance mechanism as the most prominent example, converge to PS scheduling (even in absence of losses). When the priority is exactly linear in the attained service, such as for scalable TCP and TCP's slow start phase, the shares remain constant in between file initiations and completions.

Low latency, high throughput, and fairness are the stringent performance requirements of data center networks. For web applications continue to thrive, data center must remain effective by addressing these challenges. While most of the challenges have been handled by data center specific transport protocols (e.g. DCTCP), these transport protocols are not default options for data center users. Free choice of transport protocols demands careful consideration for an effective solution. We argue that queuing scheme at layer-2 switches hold the key to address these issues. Therefore, in this paper, we present FaLL, a queuing scheme for data center networks that meets following performance requirements: fair share of bandwidth, low queuing latency, high throughput, and ease of deployment. FaLL uses an efficiency module, a fairness module and a target delay based dropping scheme to meet these goals. FaLL requires very small amount of state information. Through rigorous experiments on real testbed, we show that FaLL outperforms various peer solutions in variety of network conditions.

Denial of service (DoS) attacks are currently one of the biggest risks any organization connected to the Internet can face. Hence, the congestion handling techniques at its edge router(s), such as active queue management (AQM) schemes must consider possibilities of such attacks. Ideally, an AQM scheme should (a) ensure that each network flow gets its fair share of bandwidth, and (b) identify attack flows so that corrective actions (e.g. drop flooding traffic) can be explicitly taken against them to further mitigate the DoS attacks. This paper presents a proof-of-concept work on devising such an AQM scheme, which we name Deterministic Fair Sharing (DFS). Most of the existing AQM schemes do not achieve the above goals or have a significant room for improvement. DFS uses the concept of weighted fair share (wfs) which allows it to dynamically self-adjust the router buffer usage based on the current level of congestion, while assuring fairness among flows and aiding in identifying the malicious ones. We demonstrate the performance of DFS via extensive simulation and compare against other existing AQM techniques.

This paper describes a mechanism we call "SRED" (stabilized random early drop). Like RED (random early detection) SRED pre-emptively discards packets with a load-dependent probability when a buffer in a router in the Internet or an intranet seems congested. SRED has an additional feature that over a wide range of load levels helps it stabilize its buffer occupation at a level independent of the number of active connections. SRED does this by estimating the number of active connections or flows. This estimate is obtained without collecting or analyzing state information on individual flows. The same mechanism can be used to identify flows that may be misbehaving, i.e. are taking more than their fair share of bandwidth. Since the mechanism is statistical in nature, the next step must be to collect state information of the candidates for "misbehaving", and to analyze that information. We show that candidate rows thus identified indeed have a high posterior probability of taking a larger than average amount of bandwidth.

SUMMARY With the rapid development of wireless network technologies, heterogeneous networks with wired and wireless links are becoming common. However, the performance of TCP data transmission deteriorates significantly when a TCP connection traverses such networks, mainly because of packet losses caused by the high bit error rate of wireless links. Many solutions for this problem have been proposed in the past literature. However, most of them have various drawbacks, such as difficulties in their deployment by the wireless access network provider and end users, violation of TCP’s end-to-end principle by splitting the TCP connection, or inapplicability to IP-level encrypted traffic because the base station needs to access the TCP header. In this paper, we propose a new mechanism without such drawbacks to improve the performance of TCP over wired and wireless heterogeneous networks. Our mechanism employs a receiver-based approach, which does not need modifications to be made to the sender TCP or the base station. It uses the ACK-splitting method for increasing the congestion window size quickly in order to restrain the throughput degradation caused by packet losses due to the high bit error rate of wireless links. We evaluate the performance of our mechanism and show that our mechanism can increase throughput by up to 94% in a UMTS network. The simulation results also show that our mechanism does not significantly deteriorate