The Load Balancer will only detect a failed machine when it sends a request to the machine and it doesn't respond. If there is no traffic to your cluster, then the failed back-end will not be reported until traffic arrives.
Also, back-end machines are only marked as dead if they have failed several times. So if only one request fails, the machine will not immediately be classed as non-working. This prevents false alarms when running a service with faulty server software or poorly written CGI scripts on web sites, for instance.
You can see whether the load balancer believes each machine is alive on the GUI. Working machines are coloured grey. If they are faded out, then the balancer has not used that back-end for a while, and so its knowledge may be out of date. This is not a problem for the balancer, it will discover whether the service is working or not the next time it is used.
