Hardware Reference
In-Depth Information
500
Graph truncated: 945.7 TiB
Climate user A
All others
400
300
200
100
0
Read
Write
Read
Write
Read
Write
Small jobs
(up to 4K procs)
Medium jobs
(up to 16K procs)
Large jobs
(up to 160K procs)
FIGURE 27.3: Total amount of data read and written by Darshan-
instrumented jobs in each partition size category on Intrepid.
100 %
Used at least 1 file per process
Used MPI-IO
80 %
60 %
40 %
20 %
0 %
Small jobs
(up to 4K procs)
Medium jobs
(up to 16K procs)
Large jobs
(up to 160K procs)
FIGURE 27.4: Prevalence of key I/O characteristics in each partition size
category on Intrepid.
as having a file-per-process access pattern if it opened at least N files during
execution, where N is the number of MPI processes. Such jobs account for
31% of all core hours in the small job size category, but they do not appear
at all in the large job size category. Another job was defined as using MPI-
IO if it opened at least one file using
MPIFileopen()
. In contrast to the
file-per-process usage pattern, MPI-IO usage increases with job scale, going
from 50% for small jobs up to 96% for large jobs. The decline in file-per-
process access patterns and the increase in MPI-IO usage suggest that large-
scale applications are using more advanced I/O strategies in order to scale
effectively and simplify data management.
27.3 Conclusion
The Darshan I/O characterization tool has demonstrated that it is possible
to instrument leadership-class production applications with negligible over-
head. Since its initial development in 2009 [4], it has been in production on
