ACADEMIA

The National Center for Supercomputing Applications (NCSA) has released version 2.1 of its Virtual Machine Interface (VMI) software. The release – along with documentation, a bug reporting tool, and support mail lists -- is available at the VMI project website at http://vmi.ncsa.uiuc.edu/. VMI is middleware that addresses the issues of availability, usability, and management in the context of large-scale SANs interconnected over wide-area computational grids. With VMI, users are able to run applications on distributed clusters that use different types of interconnects to communicate among processors. The current release supports AMD Opteron, Intel X86-64, IA32, IA64, and PowerPC platforms running the Linux operating system. This release supports spanning a single MPI job across geographically distributed clusters. With the ability to load multiple communication devices at runtime, a high-performance SAN network can be used for intra-cluster communication while using TCP for inter-cluster communication transparently. An MPI communication profiling framework is included for characterizing runtime communication parameters on an individual job basis to a database for subsequent analysis. MPICH-VMI supports profile-guided optimization (PGO) at runtime for cross-cluster communication by using the profile database. Communication characteristics from previous runs are used to optimize the mapping of virtual MPI ranks to physical processors to minimize the use of high-latency, low-bandwidth links present in wide are grids. Significant scalability-related features have been added that improve start-up times for large processor counts with considerably smaller resource requirements. MPI applications have been successfully tested on up to 1,024 processors using Infiniband and 2,500 processors running Myrinet. Enhanced features for asynchronous communication have been added with this release to provide increased capability to overlap communication and computation. The RDMA API supported by VMI has been expanded to include support for RDMA Gets. VMI will transparently emulate RDMA Get semantics in an efficient manner on hardware that does not support it (Myrinet and Ethernet) or leverage hardware assisted RDMA Gets on networks that do provide the capability (Infiniband). The VMI source code is available under a liberal open source-style license that allows redistribution in source or binary form provided that copyright notices and disclaimers remain intact.