Hi,
we made the same experience but it was related to the network latency, we have always 2 round-trips (send reserve, receive response, send delete, receive response). We also use beanstalkd thru wans connected via vpn between europe, usa and asia (latency up to 200ms one way).
So we were required to build our own library which is able to have hundreds of connection (non-blocking) per worker thread, beanstalkd works fine with the amount of connections if you don't forget to increase your number of max. open files. If you want we could share our php library for handling jobs in this way…
Best,
Sebastian