The request counting is broken. Looks like synchronization problem with dirty cached reads.

Steps to reproduce:

1. start AS with some context
2. start LB
3. start 2 or more load driver threads
4. number of requests on that context goes to higher values than 2, eventually and slowly increasing

On the AS it manifests as:

19:44:14,160 WARN  [org.jboss.modcluster] (MSC service thread 1-7) MODCLUSTER000022: Failed to drain 57 remaining pending requests from default-host:/clusterbench within 10.0 seconds