INDEX
| Raweb 2003 / Project-Team : reso
|
HELP INDEX |
|
Key words: Network Quality of Service, DiffServ, PHB, RED, packet scheduling algorithm.
In the light of the frustrating experience of deployment of existing IP QoS approaches, IntServ [56] and DiffServ [55] , we have proposed a new differentiated service scheme called EDS : ``Equivalent Differentiated Services''.
This proposition represents a radical departure from traditional ``DiffServ'' architectures which rely on bounded domain concept and pricing models. The EDS is merging the Alternative Best Effort ideas [67] and the Proportional DiffServ Principles [59] . The EDS scheme aims at providing a spectrum of ``different but equivalent'' network services that offer a trade-off between delay and loss rate to the end-to-end flows. EDS acts as a network layer protocol analogous to IP and, as TCP does, the end-to-end transport layer has to do some adaptation. As EDS offers a service differentiation based on packet marking, the corresponding transport layer has to adapt data transmission and packet marking accordingly. Considering that the Internet traffic is composed of real-time traffic, interactive traffic, WEB traffic and bulk file transfer traffic, different types of adaptive packet marking algorithms, integrated in a transport protocol stub can benefit differently from the network differentiate behaviors [12] , [21] . The implementation in LINUX has been realized [44] . This software comprises different LINUX modules: a novel router mechanism merging an original RED-based active queue management algorithm and a proportional scheduling algorithm and a transport protocol for bulk data transfer that integrates an adaptive packet marking algorithm in the SCTP AIMD algorithm. This software has been functionally validated and is under performance evaluation within the European DataTAG project. The aim is to prove the EDS concept and to show that it can improve the transfer of a mix heterogeneous flows on long distance and heterogeneously provisioned links.
On a plain best-effort network, there is of course no way to control neither the end-to-end delay nor the loss rate. Packets are forwarded in a delay and with a loss probability depending on the network load. The EDS system has been designed to reflect the IP design philosophy to the plane of performance differentiation. The same way TCP has been designed to provide reliability on top of the unreliable network layer IP , we have designed three transport protocols which provide specific soft quality of service properties to applications. The RT-TP over EDS protocol ensures as best as possible an end-to-end delay and a relative reliability to a real-time application [43] . The SM-TP over EDS protocol ensures as best as possible an end-to-end delay bound to reliable short message transport [45] . The LM-TP over EDS protocol ensures as best as possible an improved end-to-end delay to bulk data transport. The NS simulation and real tests in emulated testbed demonstrate that this architecture improves flow specific performance criteria in the context of a realistic mix of heterogeneous traffics.
All the adaptive packet marking algorithms have been implemented in the SCTP NS module. The LM-TP protocol is implemented also in an SCTP module in Linux. We choose SCTP for its modular implementations both in LINUX and NS comparing to the TCP's one. The implementation as a module in Linux facilitates the test. Extensive tests with the Nistnet emulator have shown that LM-TP over EDS is resistant to non-friendly UDP flows in short or long paths. LM-TP over EDS offers a smoother and slightly better throughput than SCTP over IP. Throughput obtaines with lkSCTP, SCTP-lm and TCP has been compared in different conditions (load, delay) on different router configuration (IP default, RED, EDS, EDSRED). When there is no delay the performance of SCTP-lm on EDSRED is comparable to that of TCP. And moreover all the protocols get the best performance on EDSRED. But when the delay is set to a great value using nistnet ( RTT=200ms), the protocols get the same performance on all routers. The behavior of this protocol over EDS will be evaluated on the DataTAG link.
Performance measurement of TCP over DiffServ in production networks
Contributed by: Mathieu Goutelle, Franck Bonnassieux, Fabien Chanussot, Pascale Vicat-Blanc Primet.Key words: Network Quality of Service, DiffServ, packet marking algorithm, SLS.
End to end performance and specifically TCP behavior in real Diffserv environment have not been well experimented in large scale. To verify the end to end properties offered by Diffserv in real production networks, we have conducted a set of tests in the GEANT European backbone and in the VTHD experimental network.
The aim of these experiments was to test the behavior of TCP flow in different classes of service offered by the GEANT European backbone [41] . TCP flows marked in IP Premium, BE and LBE were analysed when LBE class, BE class and even IP Premium class are congested. For these tests, we have two PCs in NL and UK GEANT POPs connected at 1Gb/s to the core routers. Two possible paths between the PCs where created by static routing. The network bottleneck is the same for both paths (between France and Netherland via Belgium) and is 2.5Gb/s. Other links are 10Gb/s. We were also able to generate background traffic with the SmartBits in UK and DE in order to artificially congest the bottleneck FR-BE-NL. The SmartBits can generate up to 2.5Gb/s of raw traffic (not TCP). The results we obtained in GEANT show that the DiffServ implementation respects the IETF specifications [42] : IP Premium is very well protected in throughput against BE / LBE, BE is very well protected against LBE, LBE keeps the 5% of bandwidth under BE congestion. These results confirm that applications that require stable throughput can reserve IP premium resources and that unfriendly or intrusive applications that are not rate sensitive can use all the available resources without disturbing classical traffic when using LBE class.
The results we obtained in VTHD allow the evaluation of the performance of the specific DiffServ implementation of VTHD. IP Premium offers good performance stability to TCP while Assured Forwarding, enhance performance but is not able to guarantee bounds. Several performance analysis with a distributed medical images processing software have been performed [80]
Key words: Network Quality of Service, DiffServ, packet marking algorithm, SLS.
We propose a service that dynamically adaptes packet marking to best fit the requirements of individual grid flows and simultaneously to best allocate the shared differentiates resources [49] .
A Grid oriented QoS API and a programmable QoS service QoSINUS have been designed and developed within the context of the e-Toile project to introduce flexibility and dynamic in the management, the control and the achievement of end to end QoS in Grid context. Such an approach increases slightly the complexity at the Grid/WAN Network frontier points, but leaves the core network and the grid applications simple. This edge service aims at :
allowing heterogeneous Grid flows to specify individually and directly their QoS objectives,
mapping these objectives with the existing IP QoS services provided at the edge of the core inter-networks for improving the individual packet performances,
realizing a dynamic and appropriate adaptation according to the real state of the link, the QoS mechanisms configured and the experienced performances.
The first issue is addressed by an API that provides the user the ability to characterize the flow needs in terms of qualitative or quantitative end-to-end delay, end to end throughput, end-to-end loss rate or in terms of relative weight of these three main metrics. This API permits to define SLS (end to end service level specifications) in XML.
The second issue is addressed by a service architecture that combines flow aware and infrastructure aware components to map and dynamically adapt the QoS specification of the flows to the QoS facilities offered by the network. IP premium is a finite and scare resource. To avoid to waste this resource, we propose algorithms that statically or dynamically adapt the packet marking according the real QoS of a TCP flow. The analysis of ACK permits to calculate periodically the amount of data transfered and to increase packet priority when required in order to meet some deadline requirement. The ultimate goal is to provide an Earliest Deadline First algorithm in an edge packet marking equipment, in order to serve the performance requirements of individual TCP flows. This algorithm has been implemented in an active service under the TAMANOIR environment.
|
|
|