Section: Other Grants and Activities
The objective of this project is to promote the collaborations of Japan and France on grid computing technology. In order to promote the collaborative researches, we consider that this project is organized for the following three parts:
Grid interoperability and applications
Instant Grid and virtualization of grid computing resources.
RESO mainly participates to the Grid Metrics topic.
Despite the development of strong technologies in all these domains, many issues are still open about the measurement methodology itself, the emulation or simulation of Grid platforms and the understanding of Grid software stack, application performance, and fault tolerance. The Grid Metrics topics, basically gathers all researches about applications, programming models, libraries, runtimes, operating systems and network evaluation, either in synthetic environment (emulators and simulators) or real environment (real network and Grids).
AIST Grid Technology Research Center: GridNet-FJ associated team
Since 2007, RESO is pursuing its collaboration with AIST through the Gridnet-FJ associated team. We followed and even increased our working program on four parts: 1) High speed transport protocol over very high speed links, 2) Bandwidth allocation and control in Grids, 3) Optimisation of MPI communications in Grids, 4) Co-design of GtrcNET-packet capture functionality.
On point 1) with the high speed testbed for protocol evaluation we have deployed within Grid5000 and which integrates GtrcNET1 and GtrcNET10, we pursued our work on TCP variants comparison. We highlight the problem of congestion level which makes TCP behave very strangely (long TCP stops) and the problem of congesting reverse traffic. During our visit to AIST, we had long discussions on the TCP stop problem. This issue has been now solved and a patch to LINUX TCP stack has been posted; We also work together on the INRIA HSTTS (High Speed Transport Test Suite) and gathered very interesting and constructives remarks from AIST colleagues.
The collaboration between AIST GTRC team and INRIA RESO team on the point 2) aims at studying how BDTS, a scheduled data transfer service could benefit this flexibility offered by advance provisioning of some network path and to develop a service which use the interface.
AIST GTRC is collaborating with Pr Ishikawa team at University of Tokyo on GridMPI implementation (point 3). The AIST develop GridMPI, an MPI implementation designed for grids, that uses a similar system of relays as we do in our MPI5000 layer. This architecture has two goals: transmission of messages from a private cluster to a public cluster and also allowing the transfer from a Myrinet cluster to TCP on the long distance link. Their implementation is able to manage many relays to forward datas on long distance links. Their objective is different but the involved mechanisms are similar to the MPI5000 layer. The AIST planned to test their relays on Grid5000 to have a real network platform instead of emulating latency as they are doing for the moment. These experiments are realised jointly with the RESO team.
During the stay of Yuestu Kodama and Tomohiro Kudoh at ENS, the INRIA RESO team and the AIST GTRC team design and develop the Metroflux system for 10Gb/s speed and deploy 10 of such equipment within Grid5000/ALADDIN.
Collaboration with University of Lisbon, Portugal
Participant : Paulo Gonçalves.
P. Gonçalves is co-advising the PhD program of Hugo Carrão from ISEGI, University of Lisbon. A grant from the “Programme Actions Universitaires Intégrées Luso-Françaises" supports our collaboration (ends in january 2009). H. Carrão PhD defense is programmed in january 27, 2010.