Hardware Reference
In-Depth Information
Chapter 28
Iota
Mark Howison
Brown University and Lawrence Berkeley National Laboratory
Prabhat and Surendra Byna
Lawrence Berkeley National Laboratory
28.1
Features :::::::::::::::::::::::::::::::::::::::::::::::::::::::::: 317
28.2
Success Stories ::::::::::::::::::::::::::::::::::::::::::::::::::: 318
28.3
Conclusion :::::::::::::::::::::::::::::::::::::::::::::::::::::::: 320
Bibliography :::::::::::::::::::::::::::::::::::::::::::::::::::::: 321
28.1 Features
Iota is a lightweight tracing tool for diagnosing poorly performing I/O
operations to parallel file systems, especially Lustre. It collects complete traces
of POSIX I/O calls with minimal overhead, and has been tested to scale to
110,052 MPI tasks and 1.5 TB of data. Iota is freely available under a non-
commercial license from:
https://bitbucket.org/mhowison/iota
As the storage systems and the parallel file systems that service super-
computers become more complex, it is increasingly important to have finely
tuned tracing tools for investigating poor I/O performance. The POSIX I/O
layer is often informative because it is the interface between user-space code
(either the application itself or a middle-ware library like MPI-IO or HDF5)
and a kernel-space parallel file system.
Iota has similar aims to Darshan (Chapter 27, [3]), but adds two missing
features: recording of complete traces, rather than aggregates, which allows
for more flexibility in post-processed analyses; and integration with the Lustre
API to capture stripe parameters and OST assignment of files opened on
Lustre file systems.
Like Darshan, Iota traces are buffered in memory to avoid many small
writes to the trace file (in ASCII for human readability). Darshan limits the
buer's footprint by storing only aggregate values instead of a complete traces
of all I/O activity.
317
