The Cerberus Project

Technical Limitations at RIT:

Networking Efficiency:

Currently, RIT has a very impressive network. This affords students in on campus housing and almost any on campus location to connect to gigabit ethernet. For most tasks, this is entirely overkill. Saturating a 1GBe connection is difficult for personal or commodity purposes. In the scope of the project however, this internet has proven to be a limiting factor.

RIT's networking differs from an enterprise server environment where fiber internet or 10GBe interconnects are the norm. This is perhaps the largest limitation on the project. While some access to 10GBe networks is granted, it is not wide enough in scope to the project, and unrealistically expensive to institute personally. While running the thesis program, all data paths from the frame storage servers to compute nodes are saturated. This means that while the nodes and computation VM's can process images faster, they physically cannot copy the files any faster due to the limitations of the network.

Testing between select nodes on the 10GBe network shows that increasing MTU size, jumbo packets and a variety of other network optimization and running the same optimizations yields significantly faster results. This is promising in showing that migrating the project to a 10GBe or 40GBe data center environment would absolutely allow realtime processing to occur, even if currently unreachable by commodity hardware.

Networking Administration:

Due to the way in which RIT's network is structured, there are limiting factors that affect the method in which the cluster can be administered. Hostnames and IP's must be registered through RIT's Residential Computing department. The department attempts to enforce a strict limitation on machines registered to the network which is a slow process and often inhibits the addition of new nodes or machines by several days.

This also affects how VM machines are registered. If attempted elsewhere, VM's could be bridged to the network and registered independently as physical machines, which would boost program efficiency. This is not a deal breaker but does limit efficiency in the scope of the project.

Power:

Power is also another limitation for testing all nodes at one time. Currently, since all nodes have to be stored in a personally owned location (like an apartment) rather than a datacenter, power is a manner of concern. Numerous stress tests have caused power problems in RIT's University Commons apartment by tripping building breakers.

This is difficult to remedy without moving the entire setup to a datacenter, which is outside of current authorization. Machine virtualization helps to alleviate this but the system's full power requirements still far exceed what the building can supply. This is also not an issue if implemented in an enterprise environment or with more efficient hardware since the hardware being used in this test is very expensive in regards to overall energy efficacy.