As Internet services become more popular and pervasive, a critical problem that arises is managing the performance of services under extreme overload. This paper presents a set of techniques for managing overload in complex, dynamic Internet services. These techniques are based on an adaptive admission control mechanism that attempts to bound the 90th-percentile response time of requests flowing through the service. This is accomplished by internally monitoring the performance of the service, which is decomposed into a set of event-driven stages connected with request queues. By controlling the rate at which each stage admits requests, the service can perform focused overload management, for example, by filtering only those requests that lead to resource bottlenecks. We present two extensions of this basic controller that provide class-based service differentiation as well as application-specific service degradation. We evaluate these mechanisms using a complex Webbased e-mail service that is subjected to a realistic user load, as well as a simpler Web server benchmark.

This paper proposes OverQoS, an architecture for providing Internet QoS using overlay networks. OverQoS empowers third-party providers to offer enhanced network services to their customers using the notion of a controlled loss virtual link (CLVL). The CLVL abstraction bounds the loss-rate experienced by the overlay trac; OverQoS uses it to provide differential rate allocations, statistical bandwidth and loss assurances, and enables explicit-rate congestion control algorithms.

Wireless and satellite networks often have non-negligible packet corruption rates that can significantly degrade TCP performance. This is due to TCP's assumption that every packet loss is an indication of network congestion (causing TCP to reduce the transmission rate). This problem has received much attention in the literature. In this paper, we take a broad look at the problem of enhancing TCP performance under corruption losses, and include a discussion of the key issues. The main contributions of this paper are: (i) a confirmation of previous studies that show the reduction of TCP performance in the face of corruption loss, and in addition a plausible upper bound achievable with perfect knowledge of the cause of loss, (ii) a classification of the potential mitigation space, and (iii) the introduction of a promising new mitigation that employs rich cumulative information from intermediate nodes in a path to form a better congestion response.

Load balance is critical to achieving scalability for large network emulation studies, which are of compelling interest for emerging Grid, Peer to Peer, and other distributed applications and middleware. Achieving load balance in emulation is difficult because of irregular network structure and unpredictable network traffic. We formulate load balance as a graph partitioning problem and apply classical graph partitioning algorithms to it. The primary challenge in this approach is how to extract useful information from the network emulation and present it to the graph partitioning algorithms in a way that reflects the load balance requirement in the original emulation problem. Using a large-scale network emulation system called MaSSF, we explore three approaches for partitioning, based on purely static topology information (TOP), combining topology and application placement information (PLACE), and combining topology and application profile data (PROFILE). These studies show that exploiting static topology and application placement information can achieve reasonable load balance, but a profile-based approach further improves load balance for even large scale network emulation. In our experiments, PROFILE improves load balance by 50 % to 66 % and emulation time is reduced up to 50% compared to purely static topology-based approaches. 1.

This paper describes our experiences in the development of the UDP-based Data Transport (UDT) protocol, an application level transport protocol used in distributed data intensive applications. The new protocol is motivated by the emergence of wide area high-speed optical networks, in which TCP is often found to fail to utilize the abundant bandwidth. UDT demonstrates good efficiency and fairness (including RTT fairness and TCP friendliness) characteristics in high performance computing applications where a small number of bulk sources share the abundant bandwidth. It combines both rate and window control and uses bandwidth estimation to determine the control parameters automatically. This paper presents the rationale behind UDT: how UDT integrates these schemes to support high performance data transfer, why these schemes are used, and what the main issues are in the design and implementation of this high performance transport protocol.

Most standard implementations of TCP perform poorly when packets are reordered. In this paper, we propose a new version of TCP that maintains high throughput when reordering occurs and yet, when packet reordering does not occur, is friendly to other versions of TCP. The proposed TCP variant, or TCP-PR, does not rely on duplicate acknowledgments to detect a packet loss. Instead, timers are maintained to keep track of how long ago a packet was transmitted. In case the corresponding acknowledgment has not yet arrived and the elapsed time since the packet was sent is larger than a given threshold, the packet is assumed lost. Because TCP-PR does not rely on duplicate acknowledgments, packet reordering (including outof-order acknowledgments) has no effect on TCP-PR’s performance. Through extensive simulations, we show that TCP-PR performs consistently better than existing mechanisms that try to make TCP more robust to packet reordering. When the case that packets are not reordered, we verify that TCP-PR maintains the same throughput as typical implementations of TCP (specifically, TCP-SACK) and shares network resources fairly. 1

routinely fail to meet end-to-end performance expectations. The default transport control protocol for best effort data traffic is currently TCP, which does not scale well to lOOMbps and higher networks over long distances. In congestion avoidance TCP is not swift enough to fully utilize resources over paths with a high delay bandwidth product. First attempts to alleviate this problem by equipping TCP with increased aggressiveness have shown the disadvantage uf pnor fairness with the ubiquitous standard TCP-Reno, or in some cases, even amang two connections running over the same path. We propose a new delay sensitivecongestion avoidance mode (TCP-Africa) that allows for scalable, aggressive behavior in large underutilized links, yet falls hack to the more conservative TCP-Reno algorithm once links become well utilized and congestion is imminent. Through ns2 simulations we argue for the safety, efficiency, and fairness of TCP-Africa.

Abstract — In distributed heterogeneous Grid environments the protocols used to exchange bits are crucial. As researchers work hard to discover the best new protocol for the Grid, application developers struggle with ways to use these new protocols. A stable, consistent, and intuitive framework is needed to aid in the implementation and use of these protocols. While the application must not be burdened with the protocol details some of it may need to be exposed to take advantage of potential optimizations. In this paper we examine how the Globus XIO API provides this framework. We will explore the performance implications of using this abstraction layer and the benefits gained in application as well as protocol development. I.

The paper describes a general purpose high performance data transfer protocol for data intensive applications over high bandwidth networks. The protocol, named SABUL, has demonstrated the efficiency and fairness features in both experimental and practical applications. SABUL is a lightweight, reliable, application level protocol. It uses UDP to transfer data and TCP to feedback control messages. The protocol uses a rate based congestion control that tunes the inter-packet time. This algorithm is proven to be TCP friendly. In addition, the protocol also specifies the transparent memory copy avoidance.

Large-scale network simulation is an important technique for studying the dynamic behavior of networks, network protocols, and emerging classes of distributed application (e.g. Grid, peer-to-peer, etc.) Large-scale and realism are two critical requirements for network simulations of Grid application studies. Our work here extends previous efforts in three key ways. First, we study networks 100x larger than in our previous studies (20,000 routers). Second, at this scale, we study realistic network struct ures (100 AS's, BGP4 and OSPF routing) versus flat OSPF routing. Finally, we describe and evaluate a new profile-based load-balancing approach called hierarchical profile-based load balance.