LDPLFS: Improving I/O Performance without Application Modification

Abstract

Abstract-Input/Output (I/O) operations can represent a significant proportion of run-time when large scientific applications are run in parallel and at scale. In order to address the growing divergence between processing speeds and I/O performance, the Parallel Log-structured File System (PLFS) has been developed by EMC Corporation and the Los Alamos National Laboratory (LANL) to improve the performance of parallel file activities. Currently, PLFS requires the use of either (i) the FUSE Linux Kernel module; (ii) a modified MPI library with a customised ROMIO MPI-IO library; or (iii) an application rewrite to utilise the PLFS API directly.In this paper we present an alternative method of utilising PLFS in applications. This method employs a dynamic library to intercept the low-level POSIX operations and retarget them to use the equivalents offered by PLFS. We demonstrate our implementation of this approach, named LDPLFS, on a set of standard UNIX tools, as well on as a set of standard parallel I/O intensive mini-applications. The results demonstrate almost equivalent performance to a modified build of ROMIO and improvements over the FUSE-based approach. Furthermore, through our experiments we demonstrate decreased performance in PLFS when ran at scale on the Lustre file system.

Quantifying the Effects of Contention on Parallel File Systems

“…In our previous work PLFS was shown to produce a noticeable performance increase on LLNL systems under certain conditions [6], [17]. However, this paper has already demonstrated that the performance gap is reduced when using a tuned Lustre-specific MPI-IO driver.…”

“…• Finally, we re-evaluate our previous work with PLFS in order to explain the deficiencies experienced at scale [6]. By extending our contention metrics to describe PLFS-based applications, we demonstrate that at scale PLFS creates high levels of self-contention on Lustre file systems.…”

“…Although PLFS has been shown to provide increases in performance that are in-line with those of using ad_lustre with optimal configuration options, it has also been shown that at large scale, PLFS may harm performance, such that it performs worse than even an unoptimised MPI-IO installation (using the UNIX file system, POSIXcompliant driver ad_ufs) [6]. This paper analyses the potential performance of the Lustre ADIO driver over the standard UFS driver in similar way to previous studies [4], [5], [21], [22].…”

“…As we have previously demonstrated [6], [17], the large parallel file systems connected to some of the most powerful supercomputers in the world are currently being under utilised -partially due to a lack of available software drivers, partially due to lack of optimisation in the applications themselves. The work presented by Behzad et al [5] shows how using the Lustre-specific MPI-IO driver (ad_lustre) distributed with most MPI implementations can lead to performance improvements of up to 100× over the default installation.…”

A Parallel MPI I/O Solution Supported by Byte-addressable Non-volatile RAM Distributed Cache

“…Tsujita, Y. et al obtained remarkable improvements by extending Two-Phase I/O even further, by using multiple threads [7]. Other researchers are focused on new MPI I/O implementations [8] or improvements in PFS [9] [10]. The extensions of this kind, however, do not consider possibilities offered by emerging hardware technologies.…”