Extending the scope of the controlled logical clock

摘要: Event traces are helpful in understanding the performance behavior of parallel applications since they allow in-depth analysis communication and synchronization patterns. However, absence synchronized clocks on most cluster systems may render ineffective because inaccurate relative event timings misrepresent logical order lead to errors when quantifying impact certain behaviors or confuse users time-line visualization tools by showing messages flowing backward time. In our earlier work, we have developed a scalable algorithm called controlled clock that eliminates inconsistent inter-process postmortem pure MPI applications, potentially running large processor configurations. this paper, first demonstrate also proves beneficial computational grids, where single application is executed using combined power several geographically dispersed clusters. Second, present an extended version that--in addition message-passing semantics--also preserves restores shared-memory semantics, enabling correction from hybrid applications.