Tesis doctoral de Lluís FÁ brega Soler
The traditional ¿data¿ applications in the internet, such as web browsing, peer-to-peer file sharing, ftp, e-mail and others, transfer discrete messages or ¿documents¿ (a web request, a basic web file, an embedded image, an ftp file, an ftp command, etc.). They are all built on top of tcp. Documents are partitioned into blocks and sent through the network into a sequence of packets or ¿flows¿ within tcp connections. The users of these applications expect that there is no error in the transfer of documents and also that the response time is the smallest possible below a certain maximum value. Absolute fidelity is achieved through tcp¿s retransmission procedures. This packet retransmission increases the packet delay and consequently the document transfer time. Moreover, it may cause duplicated packets, which are discarded by the destination. From the point of view of the network, the decisive quality of service (qos) parameter is the average receiving rate or network throughput, which includes the duplicates. Since users expect the smallest possible response time below a certain maximum value, the network should provide the maximum possible throughput above a minimum value, a network service that we call the minimum throughput service (mts), and tcp should also be able to use it. Since the maximum possible throughput changes over time, tcp sources use rate-adaptive algorithms [jaco88a] that increase and decrease the sending rate in order to match these variations and minimize packet loss. ¿data¿ applications and tcp flows are called ¿elastic¿ due to this ability to adjust the sending rate to different network conditions. Besides a variable average rate, elastic flows are very bursty. Finally, internet traffic measurements show that the distribution of the document¿s size presents a heavy tail [crov97a], and as a consequence, most elastic flows are short and a few of them are very long. the ¿traditional¿ network scheme on the internet is based only on fifo and tail drop queues, without traffic conditioning or admission control (ac) mechanisms. The combination of this scheme with tcp rate-adaptive algorithms aims to provide a throughput equal to the fair rate of the bottleneck link. Its main advantage is its simplicity. However, it is not able to provide different throughputs to different tcp flows. Moreover, it does not provide isolation between flows, i.E., Flows sending at a higher rate than the fair throughput can damage other well-behaved flows. Finally, when resources in the followed path are enough to satisfy the minimum throughput requirements of all flows, all of them are satisfied; otherwise, i.E., During congestion situations, none of them is satisfied. an efficient way of dealing with congestion situations is to use an ac mechanism. With ac, when congestion occurs, some of the flows receive the minimum throughput (they are ¿accepted¿) and the rest do not receive it (they are ¿rejected¿ or ¿blocked¿). Congestion situations can be reduced by increasing network resources or by optimizing its use through better routing techniques. However, if congestion still occurs, ac achieves an efficient use of network resources by maximizing the number of satisfied flows. The blocking rate depends on the behavior of users¿ demands, the chosen network provisioning, the routing techniques used, and the ability of the ac mechanism to maximize the number of satisfied flows. However, using an ac mechanism complicates the network scheme, and therefore a major concern is making the ac as simple as possible. our main objective is to design a network scheme with ac for tcp elastic traffic using simple mechanisms. Specifically, a network scheme that guarantees the mts to the maximum possible number of flows, where a flow is defined as a sequence of related packets within a tcp connection, that is able to provide different minimum throughputs to different users and isolation between flows, that is built with mechanisms (of traffic conditioning, queue disciplines and ac) that reduce the per-flow state, per-flow signaling and per-flow processing as much as possible, and that considers a multidomain scenario in which each domain has a user-provider agreement with each of its neighboring domains. in this work we propose two network schemes with ac for tcp elastic traffic. Although the specific ac of each scheme is different, both schemes have a similar architecture. They are based on packet classes, with a different discarding priority assigned to each one. The ac is edge-to-edge and based on measurements, so that only edge routers participate in the ac and exchange signaling. The edge routers keep a list of active flows, which is used to detect new flows and to isolate the accepted flows by enforcing the agreed traffic profile to their input traffic. The ac is implicit: the start of a new flow is detected at the ingress when its first packet is received; the end of a flow is detected when no packet is received within a certain defined timeout interval; the request of the mts and the desired minimum throughput are indicated through the port numbers and/or a specific mark in the packets in a way that has been specified in the user-provider agreement; acceptance is indicated simply by providing the mts, while rejection is indicated by providing the best-effort service. The user-provider agreement also defines the value of the maximum aggregated throughput the user may ask for. Finally, the schemes also use pre-established logical paths from ingress points to egress points. in our 1st scheme the ac is based on per-flow throughput measurements. The first packets of the flow are used as a probing flow. Flow¿s throughput is measured at the egress, and then signaling packets carry this measurement to the ingress, where the ac decision is made. In this scheme it is necessary to use a special modification of tcp sources. The short-term fluctuations of the sending rate of a ¿standard¿ tcp source reduce the performance, and to avoid this situation, we propose a modification of tcp¿s sending algorithms that keeps the short-term sending rate close to the actual average and above a minimum value. our 2nd scheme aims at achieving better performance for the ¿standard¿ tcp sources than our 1st scheme. The ac is based on per-aggregate throughput measurements. The out-profile packets of the aggregation of accepted flows (i.E., The packets that correspond to the flows¿ extra throughput), and in second term, a special flow (with packets marked as the highest discarding priority class), are used together as a probing flow. The aggregated throughput is measured at the egress, and then signaling packets carry this measurement to the ingress, where the ac decision is made. we evaluate the performance of both schemes with simulations in several network topologies using different traffic loads consisting of tcp flows that carry files of varying sizes. We study the influence of several parameters on the performance of the schemes. The results confirm that our 1st scheme with the modified tcp and our 2nd scheme with the ¿standard¿ tcp guarantee the requested minimum throughput to the accepted flows and achieve good utilization of the network resources.
Datos académicos de la tesis doctoral «Admission control schemes for tcp elastic traffic in class-based networks«
- Título de la tesis: Admission control schemes for tcp elastic traffic in class-based networks
- Autor: Lluís FÁ brega Soler
- Universidad: Girona
- Fecha de lectura de la tesis: 11/07/2008
Dirección y tribunal
- Director de la tesis
- Teodor María Jove Lagunas
- Tribunal
- Presidente del tribunal: josep Sole pareta
- José Luis González sánchez (vocal)
- xavier Hesselbach serra (vocal)
- ramon Fabregat gesa (vocal)