Author : LUPG
Page : << Previous 4
we face when trying to debug parallel applications. The first problem is that things happen in different processes or threads at the same time, and we need a debugger that can follow all relevant runnables simultaneously. most debuggers now can handle processes properly, and on platforms with multi-threading support, usually the native debugger can handle threads as well. Of course, we need some kind of IDE if we want to use these debuggers without loosing our sanity.
However, because of the complex nature of such applications, and their sensitivity to timing issues, stopping the application in order to examine it with a debugger is something we sometimes cannot afford. Thus, our best shot would be at extensive logging, that can be turned on and off while the application is running. Such a logging facility must allow us to see which process or thread wrote each log record, and exactly when, to be able to deduce anything out of this info. We would also need to make the format of the log files easy to parse, in order to locate interesting events. It is advised that such a logging mechanism be devised early in the life of the project, as it can save hours of battling processes later.
One of the first problems we face when looking for the the cause of a bug in a parallel application, is finding the responsible process. Many times several processes are sending messages in a chain, and this chain breaks somewhere along the way. One method that could be used to debug such problems between two interacting processes, is simply suspending both of them. Then running the one that should initiate the message with a debugger, checking that the data it contains is legal and that the message it is about to send contains correct information. Then we can attach a debugger to the second process, set a breakpoint on the function that is supposed to receive the message, and resume the execution of this process. Of-course, this method cannot be employed when the message handling is sensitive to some timing constraints. In that case, only extensive logging will help us.
Page : << Previous 4