The analysis server is supposed to support multiple trace file formats. Therefore, a format unification layer is needed.
Results from earlier scalability studies for the ASCI PathForward project, revealed that reading event trace data from disk and transforming to appropriate internal data structures required a significant amount of CPU time. We therefore apply a fine-grained parallel input process that does the event translation on top of a much coarser parallel I/O process. To avoid unnecessary inter-worker communication, every worker directly accesses the trace data files via a process-local instance of a trace format support library, which allows for selective event data input. So far, we observed a satisfactory I/O performance on the tested platforms that commonly used NFS. A parallel file system however, could increase the load performance of our approach even more and fits perfectly well into the architecture.