HPC-5 Open Source Software Projects

MPI-IO Test (fs_test)

Although there are a host of existing file system and I/O test programs available, most are not designed with parallel I/O in mind and are not useful at the scale of the clusters at Los Alamos National Lab (LANL). LANL's MPI-IO Test was written with parallel I/O and scale in mind. The MPI-IO test is built on top of MPI's I/O calls and is used to gather timing information for reading from and writing to file(s) using a variety of I/O profiles; N processes writing to N files, N processes writing to one file, N processes sending data to M processes writing to M files, or N processes sending data to M processes to one file. These diagrams illustrate the various I/O access patterns. A data aggregation capability is available and the user can pass down MPI-IO, ROMIO and file system specific hints. The MPI-IO Test can be used for performance benchmarking and, in some cases, to diagnose problems with file systems or I/O networks.

The newer versions of MPI-IO test have been renamed to fs_test. The most current version of fs_test is currently available on Sourceforge.

The MPI-IO Test (fs_test) is open sourced under LA-CC-05-013.

Release	Date	Source	Documentation
1.000.21	8 July 2008	mpi_io_test_21.tgz	README
1.000.20	13 November 2007	mpi_io_test_20.tgz	README
1.000.09	15 December 2006	mpi_io_test_09.tgz	README
1.000.08	2 March 2006	mpi_io_test_08.tgz	README
Changelog
This code currently lives on Sourceforge

Rompio

Software designed to measure MPI collective, independent, asynchonous, and POSIX file I/O performance, synthesizing how applications utilize file systems in the LANL ASC computing environment. File layouts include N processors to N files and N processors to 1 file. Write access patterns include non-strided and strided. Pure network performance measures can also be obtained for relevant MPI collective calls. Refer to the README file for a detailed description of input and output.

The MPI-FTW Test is open sourced under LA-CC-06-018.

Release	Date	Source	Documentation
1.0.0	09 February 2007	rompio.v1.0.0.tar.gz	README

LANL-Trace

LANL-Trace is a tracing mechanism that integrates with MPI and uses either strace or ltrace to capture the I/O of a parallel application. In addition, LANL-Trace runs a simple MPI program both before and after the traced program in order to provide information about the possible skew and drift of the distributed clocks in order to allow researchers to meaningfully interpret the timestamps in each of the trace files of which one is produced by each process in the parallel application. More detailed information and instructions are available in the README file that comes with the LANL-Trace download.

LANL-Trace is open sourced under LA-CC-07-055.

Release	Date	Source	Documentation
1.0.0	10 November 2007	Tarball	README

LANL Checkpoint Research Tool

LANL Checkpoint Research Tool is a set of software tools designed to help researchers understand, explore, and analyze the unique patterns of IO within the checkpoint dumps of highly parallel applications run on LANL supercomputers. It contains:

Ninjat

Ninjat (formerly LANL Trace Viz) is a set of vizualization tools. One tool can be used to show an animated replay of a LANL Trace such as these. An animated tool shows which process writes or reads from which file at what offset over the lifetime of the trace. The other tool creates a static image showing the cumulative amount of I/O to each file as a function of time. Writes are shown as positive values and reads are shown as negative.

DBViz

(Formerly LANL MPI-IO Viz) A vizualization tool used to analyze experimental results from MPI-IO Test or any other tool that inserts results into a mysql server. All versions of MPI-IO Test since 1.00.020 include the ability to insert their results into a MYSQL database. DBViz queries mysql servers and allows for interactive exploration of the data with flexible vizualization options.

Parallel Log-structured File System (PLFS)

PLFS is a FUSE file-system which transparently converts N-1 patterns into N-N patterns. The user stills sees a logical view of their N-1 file but the underlying file system is presented with the much easier N-N workload.

Incremental Checkpoint

A library that applications can use to monitor which blocks of their data structures have been dirtied since the last checkpoint dump. This allows checkpoint dumps to potentially complete more quickly by sending only the dirty blocks.

LANL Trace Replay

LANL Trace Replay is an MPI program that reads a LANL Trace such as these and does MPI_Barriers wherever there are MPI Collective calls in the traces and replays the file system I/O in the traces. It can be set to sleep during idle times or it can just replay the I/O as fast as possible.

LANL Checkpoint Research Tool is open sourced under LA-CC-08-104 and is made available under a BSD license.

Release	Date	Source	Documentation
1.0	June 2010	lct.tar.gz	in tarball
LANL Trace Viz Powerpoint LANL Trace Viz Screenshot One LANL Trace Viz Screenshot Two LANL MPI-IO Viz Screenshot