DCS: Mobile Applications in Computational Grids

University Of Texas At Austin, Austin TX

Investigators

Abstract

Background Grid computing has created the opportunity for developing new classes of applications. Two such classes are applications that move between computing resources to exploit computational resources, and applications that migrate between servers to achieve resilience to hardware faults. Both of these are instances of mobile applications in computational grids. To achieve full mobility, applications must be able to move their state from one platform to another one that may have a different processor architecture (processor independence)and a different number of processors (platform-independence). The problem of converting existing static applications into mobile applications can be addressed at the language, compiler, middleware or operating systems level. An ideal solution would (i) be transparent to the application writer, (ii) permit mobility among heterogenous platforms, (iii) make applications robust with respect to hardware failures, and (iv) incur minimal performance penalty. The PI will address this problem using a combination of compiler technology and runtime systems. The solution is based on application-level checkpointing, an approach in which an application is instrumented so that it can save and restore its own state without any assistance from the operating system or architecture. To achieve the goal of mechanism transparency, compiler technology will be used to automatically instrument codes in this way. Application-level checkpointing of MPI programs requires a distributed protocol to ensure that processes are coordinated when they take their checkpoints. This protocol is executedby a thin runtime layer. The PI has implemented such a system for taking non-portable checkpoints (that is, checkpoints that can only be started on a platform identical in every way to the platform on which it was created). Studies of this system show that the overheads are small. The plan is to extend this system so that the checkpoints it takes are both architecture-independent and platform-independent. Doing this in the context of C and C++ is challenging. Experimental evaluation of the resulting system will be carried out jointly with the Pittsburgh Supercomputing Center (PSC). Intellectual Merit The wide deployment of mobile applications promises to unlock the potential of emerging grid systems. The proposed system will enable many existing parallel codes to be quickly and safely transformed into mobile applications. To demonstrate the effectiveness of our approach and systems, The PI plans to convert a number of static applications into mobile applications. These applications will be useful in their own right. Therefore, the proposed system will enable computational scientists to start using the grid without having to become grid experts. Broader Impact The tools produced by the project will be made publicly available. A number of graduate students are involved in the project - Greg Bronevetsky, Dan Marques, Milind Kulkarni, and Rohit Fernandes. Bronevetsky is an NSF graduate fellow. The project will therefore contribute to the pool of trained scientists and engineers in the US. The PI has a tradition of outreach to under-represented groups. With support from an NSF Research Infrastructure grant, The PI has hosted a number of female students from Smith College in Massachusetts, who spent the summer doing research projects. Funding from the RI project has just ended. If this proposal is funded, the PI can continue this outreach.