Letter of Intent to submit ITR Proposal PI and Co-PI Names: Ewing Lusk, University of Chicago and Argonne National Laboratory, PI Ridgway Scott, University of Chicago William Gropp, University of Chicago and Argonne National Laboratory Ralph Butler, University of North Florida Possible Participating Institutions: University of North Florida Argonne National Laboratory Possible Project Title: Towards Scalable System Software Environments Project Description: We address the problem of system software for scalable computing environments. We intend to include software components critical to application behavior, performance, and user convenience on large-scale parallel computers, but not to include the applications themselves nor services routinely provided by single-computer operating systems. We include as system software message-passing libraries, job schedulers, process managers, and performance analyzers. We are particularly interested in developing standard interfaces among these components. Recent developments in high-end computing lead one to expect that for some time to come large-scale computers will consist of networked shared-memory multiprocessors. Much of the system software needed by applications in such an environment already exists. Two problems, however, pose serious obstacles. The first is the scalability of the components. Systems that have been highly successful on dozens of machines need new internal algorithms and data structures to be scalable to thousands of processors. The second problem is the integration of the components. Existing machine manager daemons, process control daemons, and message-passing support systems are incompatible. The answer is not a monolithic, single integrated environment, but rather a set of publicly accepted interfaces so that components of the environment can be separately developed but still work together. We intend to conduct the necessary computer science research to address some of the more pressing problems in these areas. Specific topics will include: MPI implementation performance on SMP clusters. This will involve integration of shared-memory communication with the sophisticated use of TCP and other networks (VIA, Myrinet, etc.) at the MPI implementation level, together with high-level constructs for application-directed aggregation of messages. A prototype component-based job manager/process manager that can interact with multiple job schedulers and interface to multiple message-passing libraries. We will identify, with an eye to eventual standardization, the interfaces among various components of the run time environment that will allow multiple schedulers, process managers, I/O systems, etc. to be separately developed yet work together. A novel aspect of this prototype is our intention to verify it. We will use program verification techniques together with the Boyer-Moore ACL-2 system to prove theorems concerning the lack of race conditions in our prototype. Scalable versions of Unix commands have been specified in the literature (by us), but no robust implementations yet exist. We will explore whether these tools can now be implemented in MPI, with a specified set of auxiliary functions provided by the runtime environment. The main point is to identify the interfaces to other system software components necessary to support this functionality. The new large Linux cluster at Argonne National Laboratory, which as been acquired specifically to support computer science research on scalable system software, will allow us to test the results of our research in a relatively large-scale environment. Submitting Institution: The University of Chicago Ewing Lusk The Computation Institute and Department of Computer Science The University of Chicago 100 E. 58th Street University of Chicago 5640 S. Ellis Avenue Chicago, IL 60637 Phone: (630) 252-7852 FAX (630) 252-5986 Email: lusk@sphere.uchicago.edu