Fast and efficient input/output is a very important aspect of today’s high performance computing. Often these systems are very complex so no prediction can be made what an I/O call at the program level will trigger at the system level. One kind of such complex systems are parallel programs based on MPICH2 for message passing and PVFS2 as parallel file system. This thesis is about enhancing the environment in a way that we can find the necessary information in the traces of a program run. MPI-IO calls can now be related to disk operations. The concept of our extension of the involved software packages is to generate a unique ID for every MPI-IO call and to pass it along to the I/O-servers. After they got written into the corresponding traces we can find matching events in them and visualize them in the Jumpshot trace viewer. An easy to follow example is presented and detailed hints for future work are given.
|Item Type:||Bachelor thesis|
|Faculties / Institutes:||The Faculty of Mathematics and Computer Science > Department of Computer Science|
|Controlled Keywords:||Supercomputer, Hochleistungsrechnen, LINUX, Datenspeicherung, Ein-Ausgabe, MPI, Verteiltes Dateiverwaltungssystem|
|Uncontrolled Keywords:||Paralleles Dateisystem , PVFS2 , Leistungsanalyse , SpurerzeugungParallel File System , PFVS2 , Performance Analysis , Trace Generation|