Why do we build data center clusters the way we build them? We started this project in 2008 wondering if there was a better hardware platform with a co-designed software stack that would make it easier/more preformat to write the sort of applications that run in data centers. We wanted to try and address the many issues, which we largely saw as being a function of the hard separation between the network and the processing elements of a clusters, applications had to treat the network as a black box and then work around the issues.
We took ideas from distributed systems, networking and High Performance Computing (HPC) to create the CamCube platform, and the software stack that we now call CamCubeOS. We have demonstrated the potential performance gains for many applications, including a MapReduce-like framework (CamDoop). The CamCube platform used a 3D Torus topology (frequently used in the HPC community), for several reasons, including the fact that the physical topology matched closely the virtual topology of the CAN structured overlay, which then allowed us to borrow ideas used with structured overlays, to make designing distributed services easier.
We had to make a number of compromises in CamCube, as we needed to build an experimental system and did not want to design custom hardware or ASICs. However, in the last couple of years we have seem many exciting advancements in Industry in the CamCube direction, in particular the development and deployment of in-rack fabrics that closely integrate the network and CPUs using direct connect networks, effectively embedded the switch across all the servers. The AMD SeaMicro platform that is targeted at commodity data centers and uses a 3D Torus topology, and platforms like the Boston Viridis which uses Calxeda’s EnergyCore processors that support several topologies.
As we begin to see these systems deployed there is still the important question of how visible to the applications running on these racks will be the underlying network/fabric topology? At the moment there is a trend to try and hide the topology, and make it look and feel like a traditional ToR Ethernet switch based network. One of the many lessons from Project CamCube is that there is value in not treating the network as a black box.