Profiling Of Rt-Picls Code

It was observed, that the RT-PICLS code ran by FWKT on the hypnos cluster was producing an unusual amount of system load, according to Ganglia metrics. Since this may point to an IO-problem in the code, this code was analyzed more closely.

RT-PICLS was run with additionally provided input data on 64 CPU cores (AMD Opteron) in a single node of the hypnos cluster. Score-P 4.0 with IO-tracing support was used for profiling and tracing of the application. A developer version of Vampir 9.2 [3] enabled graphical analysis of the traces.

Different file systems were evaluated for storing the output:

<strong>bigdata</strong>: the high-performance file system of the cluster based on GPFS.
<strong>nfs</strong>: a loop-mounted EXT4 image mounted to the current compute node via NFS.
<strong>scr</strong>: the local scratch file-system of the compute node.
<strong>tmpfs</strong>: a virtual filesystem, physically located in the main memory of the compute node.

The trace files that were used to compile the referenced report are collected in this upload.