Fast Parallel Non-Contiguous File Access

Abstract

Many applications of parallel I/O perform non-contiguous file accesses: instead of accessing a single (large) block of data in a file, a number of (smaller) blocks of data scattered throughout the file needs to be accessed in each logical I/O operation. However, only few file system interfaces directly support this kind of non-contiguous file access. In contrast, the most commonly used parallel programming interface, MPI, incorporates a flexible model of parallel I/O through its MPI-IO interface. With MPI-IO, arbitrary non-contiguous file accesses are supported in a uniform fashion by the use of derived MPI datatypes set up by the user to reflect the desired I/O pattern.Despite a considerable amount of recent work in this area, current MPI-IO implementations suffer from low performance of such non-contiguous accesses when compared to the performance of the storage system for contiguous accesses. In this paper we analyze an important bottleneck in the efficient handling of non-contiguous access patterns in current implementations of MPI-IO. We present a new technique, termed listless I/O, that can be incorporated into MPI-IO implementations like the well-known ROMIO implementation, and completely eliminates this bottleneck. We have implemented the technique in MPI/SX, the MPI implementation for the NEC SX-series of parallel vector computers. Results with a synthetic benchmark and an application kernel show that listless I/O is able to increase the bandwidth for non-contiguous file access by sometimes more than a factor of 500 when compared to the traditional approach.

Exploiting Shared Memory to Improve Parallel I/O Performance

“…Datatype iterators are also similar to the flattening-on-the-fly technique (like dataloops but with added optimizations useful for vector processors) of listless I/O, which specializes the MPI pack/unpack interfaces to perform partial processing [9]. However, our datatype iterator interface appears to be unique: it allows a data transfer where both source and destination buffers are non-contiguous, and factors out separate seek and tell operations while still supporting partial processing.…”

aIOLi: an input/output library for cluster of SMP

“…6 Related works As mentioned in section 2, lot of works have been carried out on the parallel I/O issues but none fitted our needs. Recent [32] has shown some improvements by using newer MPI internal calls (flatening-on-the-fly) which generate lesser overhead. In [12], a several routines to control in a better way the mapping between file layout and physical placement by using the database concept of "view" is described.…”

The Impact of File Systems on MPI-IO Scalability

“…Several groups have implemented support for efficient noncontiguous I/O, including listless I/O [15], data sieving [14], list I/O [3], and datatype I/O [2].…”

“…-Efficient noncontiguous I/O support -Consistency semantics closely matching the MPI-IO model -Client-independent references to files Efficient noncontiguous I/O support in the file system has been a focus of a great deal of recent research [15]. The datatype I/O concept [3] in particular provides an efficient infrastructure for noncontiguous I/O in MPI-IO, with similar concepts seen in the View I/O [7] work.…”