A method for monitoring a plurality of servers in a cluster and taking corrective action for the servers. A request to one of the servers is sent. Then, a determination is made if the one server successfully handles the request and how long it took for the one server to handle the request. If a response is received indicating that the one server successfully handled the request, but it took the one server longer than a predetermined time period to handle the request, a dispatcher for the one server is notified to reduce, but not eliminate, a workload of the one server. There is specified a number of consecutive requests that can be sent to a server and not handled by the server within a specified time period for each of the requests; the number indicates that the server is down. A request is sent to one of the servers, and a determination is made that the one server did not successfully handle the request within the specified time period. A determination is made that the number has not yet been attained and therefore, no corrective action is taken. A subsequent request is sent to the one server, and a determination is made that the one server did not successfully handle the request within the specified time period. A determination is made that the number has been attained and therefore, corrective action is taken.