D. Kranzlmüller (Austria) and L. Lefèvre (France)
nondeterminism, program analysis, record&replay, moni toring.
Nondeterministic program behavior leads to different re sults in successive program executions, even if the same input data is provided. For this reason, re-executions of a program (as needed during cyclic debugging) are only pos sible, if certain precautions are taken. The most common solution is provided by record&replay mechanisms, where an initial record phase is used to extract characteristic be havioral data, which is afterwards used to control equiva lent executions during subsequent replay phases. With the novel record&replay mechanism on Myrinet network inter face cards (NIC)1 , program perturbations during the record phase are avoided by performing the initial monitoring ac tivities directly on the NIC. This approach ensures, that the CPU of the computing nodes is not affected by the mon itoring activities, while subsequent re-executions can still be controlled with the data collected on the NICs.
Important Links:
Go Back
