It is the opinion of the author that the leveling layer of a TacoPleX cluster will very seldomly be the bottleneck of a system. The time it takes to pass on a packet is _very_ small compared to the time it takes to handle a service request. The servers will in almost every case be the rate-determining-step. One can envision a situation where the server to leveler ratio is very high and where the leveling layer becomes the slow point in the system. This ratio needs to be determined.
It is unproven that having a dedicated return path in the network setup will yield a worthwhile performance increase. A cluster with a "back door" router needs to be tested against one without.
In the current system, the levelers send arp messages to get the attention of the input router. This results in chunky and uneven leveling across the levelers. A more even distribution could be obtained if leveling decisions were made at the level of the router. The problem with this system is that the router would have to be informed of the addition or removal of a router. This might be achieved via a standard routing protocol.
TacoPleX is a project is search of a maintainer. It was developed for proprietary use but never put into action and is no longer being developed. If you are interested in taking over this project, please contact the project admins.