Trace Data to Support and Enable Computer Science Research

In order to enable open computer science research access to trace data is needed. The following sets of data are provided under universal release to any computer science researcher to use to enable computer science work.

All we ask is that if you use these data in your research that you recognize Los Alamos National Laboratory for providing these data.

MPI_IO_TEST traces

These traces were collected using LANL-Trace (V 1.0.0) on the LANL MPI_IO test (V 1.00.020) application. These traces are all from system data machine number 25 on this computer systems table.

Here is the README that explains how LANL-Trace works and what the output files look like: TRACE README.

In terms of overhead, LANL-Trace does impose a fair amount of overhead. For these traces, this overhead is especially severe for small block sizes. This is intuitive since each run does a comparable amount of total I/O. Small blocksizes result in more traced events whereas large blocksizes do the same amount of total I/O using many fewer traced events. Here are graphs showing the overhead imposed on the effective write bandwidth (i.e. including open and close times) of the runs.

Note about the N-N traces. Due to a 32-byte default limitation in the length of strings, the N-N traces appear to actually be accessing a single shared file. Actually, they are accessing N different files but the 32-byte prefix of these files is the same. When the distinct filenames were truncated to 32 bytes, they appear to be the same. We are working to redo these traces with the appropriate option to ltrace to remove this limitation; in the meantime, any analysis of the N-N traces should manually alter the filenames to make them distinct. Sorry for any inconvenience.

N-to-N

N-to-1 nonstrided

N-to-1 strided

To save yourself some clicking, here is a link for all traces in a single tarball: One big tarball.

Please email us with any questions about these traces.

Previously released N-N traces

We had previously released several N-N traces. However, since then, we have had several components of our storage stack increment versions. We then acquired our N-1 traces with the newer storage stack. In order to have consistent traces, we have redone the N-N traces (which are available above). However, anyone who downloaded the N-N traces before 2008 will have downloaded the older traces. They are archived here.