Abstract

1 Introduction Often, processes must acquire exclusive access to a range of bytes. One appli-cation of byte-range locks is to implement the atomic mode of access defined in MPI-IO, the I/O interface that is part of MPI-2 [7]. MPI-IO, by default,supports weak consistency semantics in which the outcome of concurrent overlapping writes from multiple processes to a common file is undefined. The user,however, can optionally select stronger consistency semantics on a per file basis by calling the function MPI File set atomicity with flag=true. In this mode,called the atomic mode, if two processes associated with the same open file write concurrently to overlapping regions of the file, the result is the data written byeither one process or the other, and nothing in between. In order to implement the atomic mode, either the underlying file systemmust provide functions that guarantee atomicity, or the MPI-IO implementation must ensure that a process has exclusive access to the portion of the file itneeds to access [13]. Many POSIX-compatible file systems support atomicity for contiguous reads and writes, such as those issued by a single read or writefunction call, but some high-performance parallel file systems, such as PVFS [1]

Citations