Hugepage Library






High-speed Network Interface with Collective Operation Support for Cell BE (NICOLL)

Project Description (Abstract)

The Cell's heterogeneous multi-core implementation including its three level software-controlled memory hierarchy has tremendous potential for HPC environments. Clustering Cell's for performance aggregation requires low-latency and high-bandwidth network interface controllers (NIC). Moreover, for extreme scaling NICs are needed that support collective communication operations highly efficient. Whereas message-passing style communication has well proven its potential in massive-parallel environments there is still an ongoing research on how to couple NICs to fit best the HW-SW infrastructure of multi-core processing nodes. Competing concepts of protocol offloading have to be revaluated in heterogeneous multi-core architectures like Cell BE, especially if Cell is deployed as an accelerator to conventional processors in hybrid systems.

The objective of this proposal is to develop a prototype solution for an interface architecture that fits best the needs of Cell's HW-SW infrastructure in hybrid systems. If necessary, techniques of device virtualization will be applied. The focus is on interfacing InfiniBand as an example solution for a high-speed cluster interconnect technology. With respect to collective communication operations the focus is on the Message-passing Interface (MPI).

Aspects of how to couple InfiniBand, Routing and MPI acceleration cores will be studied with a special focus on computational collectives. The concepts will be validated by a prototype test hardware which consists of a Cell BE that is interfaced via FPGA to the network. The preferred coupling point of FPGA to Cell is Cell's memory controller interface for tight integration. The coupling to the conventional processor will be established using a second FPGA interface. Hereby HyperTransport (Torrenza) is the preferred interconnect technology as we envisage AMD Opteron as main processor. The software stack needs to be developed in close co-operation with the hardware development, thus enabling mutual enhancement of the designs.


We are searching for student assistants!

This work is part of a collaboration with the Center of Advanced Study IBM Böblingen, Development and Research, Germany.

Please mail to Prof. W. Rehm rehm@informatik.tu-chemnitz.de