INDEX
| Raweb 2003 / Project-Team : reso
|
HELP INDEX |
|
RESO is focusing on communication software, services and protocols in the context of High Performance Networking and applying its results to the domain of Grids.
RESO has been an INRIA pre-project proposed in 1999 between INRIA and Université Claude Bernard of Lyon.
The team joined the ``Laboratoire de l'Informatique du Parallélisme'' (LIP) - Unité Mixte de Recherche (UMR) CNRS-INRIA-ENS with Université Claude Bernard of Lyonin January 2003.
RESO has presented a project proposition to the Rhône-Alpes Research Unit Project Committee in March 2003. The INRIA RESO project has been officially created the 1st of December 2003.
The RESO approach relies on the analysis of limitations encountered in existing systems and on the theoretical and experimental exploration of new approaches. This research framework between a new specific application context and challenging network context, induces a close interaction with the application level and with the underlying network level. The methodology is based on a study of the high end and original requirements and on experimental evaluation of the functionalities and performances of high speed infrastructures. RESO gather expertises in advanced high performance local area networks protocols, in distributed systems and in long distance networking. This background work provides the context model for innovative and adequate protocols and software design and evaluation. Moreover, the propositions are implemented and experimented on real or emulated local or wide area testbeds with real conditions and large scale applications.
RESO aims at providing software solutions for high performance and flexible communications in very high speed wired networks. Current communication software and protocols designed for standard networks and traditional usages expose strong limitations when applied to this context of high performance computing. The goal of our research is optimization and control of the end to end quality of service in high performance distributed systems called computational grids.
RESO creates open source code, distributes it to the research community for evaluation and usage. The long term goal is also to contribute to the evolution of protocols and networking equipments and to the dissemination of new approaches.
The various research areas cover high performance communication software design and optimization for end systems or cluster, enhancement of network IP layer with service differentiation or active service processing, unicast and multicast transport protocols for long distance communication, network performance measurement and monitoring, real or emulated testbed design, deployment and performance evaluation.
RESO applies its research to the domains of high performance computing and to Grid communications. In a Grid, the network performance requirements are very high and may strongly influence the performance of the whole distributed system. As Grid applications generally rely on a complex interconnection of heterogeneous IP domains, end-to-end flow performances cannot be guaranteed or predicted. Thus, for achieving end-to-end QoS objectives, the remaining deficiencies of the network performances have generally to be masked by adaptation performed at the host level. RESO designs services and software to avoid the applications to be network-aware and to simplify the programming and to optimize the execution of their communication parts while fully exploiting the capacities of the network infrastructure.
During this year, RESO had main contributions in the following fields:
Protocols and optimized software for High performance PC-cluster networks;
End to End Service differentiation in IP networks;
High Performance transport protocols;
High Performance Active Networks and Services;
Network Grid services and applications.
Design and proposition of a new networking subsystem architecture built around a packet classifier executed in the Network Interface Controller (NIC). Development of the KNET system;
Study and design of efficient remote data access for clusters, that maximizes the underlying network utilization. Development of the ORFA (Optimized Remote File-system Access) software protoype on Myrinet networks.
Proposition and design of an alternative solution for end to end service differentiation in TCP/IP environment, named equivalent differentiated services(EDS). This proposition aims at enhancing performance differentiation at the IP level without requiring any control plane. Different end-to-end packet marking protocols have been designed and evaluated to prove the validity of the EDS incremental and soft evolution of the IP forwarding paradigm. EDS suite has been developed as a new queuing discipline and a modified SCTP module in LINUX kernel.
Design of a control service that dynamically manages the Diffserv classes allocated to an access point [49] . This control service attempts to optimize the utilization rate while honoring the requirement of individual flows. This service is implemented as a QoS API and an active service of the Tamanoir architecture.
Evaluation and analysis of the real behavior of TCP with IP Premium, Assured Forwarding and Less Than Best Effort DiffServ classes, in the context of the European GEANT backbone in collaboration with the DANTE consortium and in the VTHD national network. The aim was to measure the benefit that Grid flows can expect from such advanced network services and to verify the properties of the QoS services in a production environment.
Exploration of innovative approaches based on a better knowledge and light weight control of the path to solve the problem of high and controlled throughput in very high speed links with long latency. The contributions in this area are:
Analysis of limitation of existing or proposed high performance transport protocol's design and implementation in very high performance environments;
Proposition of a new approach of congestion control in this context, based on back-pressure flow control;
Analysis of on-intrusive methods of throughput measurement and proposition of an original hop by hop method for link capacity estimation and path utilization rate evaluation. A tool implementing this method, TraceRate has been developed.
We conducted several research investigations on the topic of programmable and active networks and services for Grid support :
Development of a high performance active network architecture (Tamanoir) and associated tools (Echidna, Pangolin). Proposition of load balancing functions in cluster-based active routers.
Validation of Tamanoir through internal and external projects (IBP, deployment of FPTP (LAAS, Toulouse), deployment of collaborative web caches (INSA, Lyon));
Design and development of high performance active services (DYRAM, QoSINUS);
Deployment of active and programmable solutions around VTHD backbone (RNRT VTHD++ project) and to support Grid applications (RNTL e-Toile).
We conducted several analyses on requirements and experiments on network services in the context of large scale grid projects.
Important contribution to the design and development of the eToile, national Grid testbed.
Contribution to the development and improvement of high-level end-to-end performance measurement network services for grid environment. In particular we define, develop and deploy a Probes coordination protocol PCP that aims to coordinate the concurrent measurements of this distributed system in a grid network.
Design and development of a framework for Network Cost Estimation Service, NCES and evaluation of its pertinence and accuracy in a real Grid (DataGRID) for data replica optimization.
|
|
|