In explicit or delay-driven congestion control, a common objective is to sustain high throughput without long queues and large losses at the bottleneck link of the network path. Congestion control protocols strive to achieve this goal by transmitting smoothly in the steady state. The discovery of the appropriate steady-state transmission rates is a challenging task in itself and typically introduces additional queuing and losses. Seeking insights into the steady-state profiles of queuing and loss achievable by real protocols, this paper presents an AIST (Asynchronous arrivals with Ideally Smooth Transmission) model that abstracts away transient queuing and losses related to discovering the path capacity and redistributing it fairly among the packet flows on the bottleneck link. In AIST, the flows arrive asynchronously but transmit their packets at the same constant rate in the steady state. For the link with an overprovisioned buffer, our queuing-theoretic analysis and simulations for different smooth distributions of packet interarrival times agree that queuing under AIST with the target utilization of 1 is on the order of the square root of N, where N is the number of flows. With small buffers, our simulations of AIST show an ability to provide bounded loss rates regardless of the number of flows.

Data center networks are characterized by high link speeds, low propagation delays, small switch buffers, and temporally clustered arrivals of many concurrent TCP flows fulfilling data transfer requests. However, the combination of these features can lead to transient buffer overflow and bursty packet losses, which in turn lead to TCP retransmission timeouts that degrade the performance of short-lived flows. This so-called TCP-incast problem can cause TCP throughput collapse. In this paper, we explore an application-level approach for solving this problem. The key idea of our solution is to coordinate the scheduling of short-lived TCP flows so that no data loss happens. We develop a mathematical model of lossless data transmission, and estimate the maximum good put achievable in data center networks. The results indicate non-monotonic good put that is highly sensitive to specific parameter configurations in the data center network. We validate our model using ns-2 network simulations, which show good correspondence with the theoretical results.

We explored the TCP-incast throughput collapse problem in data center networks from an application-layer perspective. In particular, we presented the model and analyzed the performance of the application-based approach under TCP-incast scenario. The main idea of the approach is to schedule the server responses to data requests so that no packet losses occur at the bottleneck link. The main result we derive is the achievable goodput of a data center application if using lossless scheduling. The simulations confirmed the validity of our model and the results derived through the theoretical analysis. Our future work is to explore practical scenarios of our approach and implement it in a real data center network.

The single best-effort service of the Internet struggles to accommodate divergent needs of different distributed applications. Numerous alternative network architectures have been proposed to offer diversified network services. These innovative solutions failed to gain wide deployment primarily due to economic and legacy issues rather than technical shortcomings. Our paper presents a new simple paradigm for network service differentiation that accounts explicitly for the multiplicity of Internet service providers and users as well as their economic interests in environments with partly deployed new services. Our key idea is to base the service differentiation on performance itself, rather than price. We design RD (Rate-Delay) network services that give a user an opportunity to choose between a higher transmission rate or low queuing delay at a congested network link. To support the two services, an RD router maintains two queues per output link and achieves the intended rate-delay differentiation through simple link scheduling and dynamic buffer sizing. We evaluate the performance, deployment, and security properties of the RD network services in various network topologies and traffic scenarios including delay-sensitive VoIP (Voice over Internet Protocol) applications.

While Transmission Control Protocol (TCP) variants with delay-based congestion control (e.g., TCP Vegas) provide low queueing delay and low packet loss, the key problem with their deployment on the Internet is their relative performance when competing with traditional TCP variants with loss-based congestion control (e.g., TCP NewReno). In particular, the more aggressive loss-based flows tend to dominate link buffer usage and degrade the throughput of delay-based flows. In this paper, we study a novel approach for achieving fair sharing of the network resources among TCP variants, using Rate-Delay (RD) Network Services. In particular, loss-based and delay-based flows are isolated from each other and served via different queues. Using extensive ns-2 network simulation experiments, we show that our approach is effective in providing fairness between loss-based NewReno and delay-based Vegas flows.

With the Internet offering a single best-effort service, there have been numerous proposals of diversified network services that align better with the divergent needs of different distributed applications. The failure of these innovative architectures to gain wide deployment is primarily due to economic and legacy issues, rather than technical shortcomings. We propose a new paradigm for network service differentiation where design principles account explicitly for the multiplicity of Internet service providers and users as well as their economic interests in environments with partly deployed new services. Our key idea is to base the service differentiation on performance itself, rather than price. The proposed RD (Rate-Delay) services enable a user to choose between a higher transmission rate or low queuing delay at a congested network link. An RD router supports the two services by maintaining two queues per output link and achieves the intended rate-delay differentiation through simple link scheduling and dynamic buffer sizing. After analytically deriving specific rules for RD router operation, we conduct extensive simulations that confirm effectiveness of the RD services geared for incremental deployment in the Internet.

In this paper, we present a model for investigating lower bounds on queuing under smooth congestion control with overprovisioned buffers. We consider an idealized protocol where all flows always transmit at equal rates. The ideally smooth transmission does not eliminate queuing altogether because packets of different flows might arrive to a node simultaneously due to asynchronous arrivals of the flows, which is intrinsic to computer networks. A prominent feature of our model is its simplicity, making analysis tractable and experiments scalable. Our results reveal steady-state queues of at least O(radicN) packets, where N is the number of flows. Hence, no congestion control protocol is able to avoid losses at a fully utilized link with a constant buffer shared by arbitrarily many flows.

To discover an efficient fair sending rate for a flow, Transmission Control Protocol (TCP) saturates the bottleneck link and its buffer until the router discards a packet. Such TCP-caused queuing is detrimental for interactive and other delay-sensitive applications. In this paper, we present Multimodal Control Protocol (MCP) which strives to maintain low queues and avoid congestion losses at network links. The multimodal MCP engages routers and hosts in limited explicit communication. A distinguishing property of MCP is stable transmission after converging to efficient fair states. To ensure convergence to fairness, MCP incorporates an innovative mechanism that enables a flow to urge all flows sharing its bottleneck links to operate in a fairing mode, dedicated to fairness improvement. To make the stable fair rates independent of round-trip times and packet sizes, MCP employs rate-based control and uniform timing of adjustments.

Interactive and other delay-sensitive applications are interested in keeping end-to-end delays of their packets minimal. Unfortunately, congestion control offered by transmission control protocol (TCP) and other existing protocols inflates the end-to-end delays by building up queues at bottleneck links. In this paper, we investigate multimodal control protocol (MCP) designed to maintain low queues after converging to the stable state where MCP flows utilize shared bottleneck links efficiently and fairly. To achieve this goal, MCP incorporates multiple modes of operation and allocates few bits in each packet header for explicit communication between hosts and routers. An innovative aspect of the explicit communication mechanism is an ability of a flow to urge all flows on its bottleneck links to switch temporarily into a fairing mode and thereby improve fairness of the bottleneck sharing. To make the fair sharing independent of round-trip times and packet sizes, MCP uses the sending bitrate as a control parameter and employs uniform adjustment timing for all flows. Our evaluation of MCP demonstrates its efficient fair operation and significantly shorter stable-state queues than under existing congestion control protocols.