Hardware Reference
In-Depth Information
FIGURE 3.6: Snap-shot of the 10 largest running jobs on Blue Waters at
February 11, 2014. At this time 99.8% of the nodes were in use by scientific
jobs. [Image courtesy of Dave Semenaro.]
3.4.1 Science and Engineering Team Application I/O
Requirements
For a majority of applications, I/O performance was identified as crucial
for computational performance for scalable programs. Quantified estimates of
application input (initial condition or restart), checkpoint and output (visual-
ization, final state) file sizes, and counts were gathered from many of the SETs.
Input volumes ranged from a single 100-GB file for molecular dynamics prob-
lems to 10 TB (across tens of thousands of files) for turbulence applications
to 75 TB (across hundreds of thousands of files) for geophysical applications.
Checkpoint and output sizes had comparable ranges. Defensive checkpoint in-
tervals will vary according to the observed mean time to interruptions and
node counts used for the jobs.
Estimates for total storage requirements for SETs' campaigns ranged from
10 TB to over 10 PB. The survey also highlighted the variety of application
I/O schemes in use ranging from \le-per-process" to dedicated I/O writer
tasks embedded in the applications, to \many-to-one" implementations for
storage. Nine SETs utilize high-level I/O libraries such as HDF5; three SETs
use NetCDF; and other SETs, such as Lattice QCD, which use USQCD's par-
allel QIO package and BoxLib from LBNL. For the remaining teams, POSIX
I/O remains a popular choice.
In practice on Blue Waters, all the methodologies are working really well.
None of the teams have reported insucient I/O because Blue Waters is
