Participants : Serge Petiton [correspondant] , Nahid Emad, Maxime Hugues.
Scientific end-users face difficulties to program P2P large scale applications using low level languages and middleware. We provide a high level language and a set of tools designed to develop and execute large coarse grain applications on peer-to-peer systems. Thus, we introduced, developed and experimented the YML for parallel programming on P2P architectures. This work was done in collaboration with the PRiSM laboratory (team of Nahid Emad).
The main contribution of YML is its high level language for scientific end-users to develop parallel programs for P2P platforms. This language integrates two different aspects. The first aspect is a component description language. The second aspect allows to link components together. A coordination language called YvetteML can express graphs of components which represent applications for peer-to-peer systems.
Moreover, we designed a framework to take advantage of the YML language. It is based on two component catalogues and an YML engine. The first one concerns end-user's components and the second one is related to middleware criteria. This separation enhances portability of applications and permits real time optimizations. Currently we provide support for the XtremWeb Peer-to-Peer middleware and the OmniRPC grid system. The support for Condor is currently under development and a beta-release will be delivered soon (in this release, we plan to propagate semantic data from the end-users to the middleware). The next development of YML concerns the implementation of a multi-backend scheduler. Therefore, YML will be able to schedule at runtime computing tasks to any global computing platform using any of the targeted middleware.
We experimented YML with basic linear algebra methods on a XtremWeb P2P platform deployed between France and Japan. Recently, we have implemented complex iterative restarted Krylov methods, such as Lanczos-Bisection, GMRES and MERAM methods, using YML with the OmniRPC back-end. The experiments are performed either on the Grid5000 testbed of on a Network of Workstations deployed between Lille, Versailles and Tsukuba in Japan. Demos was proposed on these testbeds from conferences in USA. We recently finished evaluations of the overhead generated using YML, without smart schedulers and with extrapolations due to the lack of smart scheduling strategies inside targeted middleware.
In the context of the FP3C project funded by ANR-JST, we have recently extended YML to support a directive distributed parallel language, XcalableMP http://www.xcalablemp.org/ . This extension is based on the support of the XcalableMP language inside YML components. This allows to develop parallel programs with two programming paradigm and thus two parallelism levels. This work is a part of the project that targets post-Petascale supercomputer that would be composed of heterogeneous and massively parallel hardware.
The software is available at http://yml.prism.uvsq.fr/