We've recently begun executing a test which is very transaction heavy, meaning each user performs alot of activity within the system before logging out and starting again. In our current environment we have run tests with 125 concurrent users successfully performing less transaction heavy tests using 3 agents. However now a 62 concurrent user version of the new test is consistently producing user failures at the end of the test on several of the agents. Originally with the 3 agents typically 20 of the users would fail. When we added another 3 agents to the mix the errors were reduced to 6 or so. It appears the controller is unable to retrieve results for the failed users, although it retrieves them for the successful users. The errors generated are too generic to be helpful. Would you think a memory upgrade on the Agent machines would do the trick? Does anyone have recommended specs for an agent machine? Currently most of the agents are at 256k of memory.
This could also be a data base error. The amount of data stored for the errors in the results repository could be more than the system can hold. Try removing old data from your results repository and see what happens.
I'm not sure of the type of application you are testing but the agents I use are quite hefty.. i.e. dual processor NT boxes with 1 Gb of ram in each box. No problems with user failures. or failing that I know that silkperformer will flag up errors if the target server has an NFS share.