Efficient protocols

The increasing demand for more and more computing power causes steady advancements of High Performance Computing (HPC) systems. The more powerful these systems will be in the future the further the number of processing units increases. A particularly important point in this context is the latency of the communication among those units, which significantly increases by the distance between two communication partners. One approach to positively influence the latency behavior is optimizing the underlying protocol structures in the overall system. Nowadays, different protocols are used for different communication distances. The latency can be improved by changing the protocol structure with two approaches. On the one hand, the used protocols can be changed to optimize the latency. On the other hand, the protocol structure can be unified. Thus, time-consuming protocol translations can be eliminated. In order to achieve this, a completely new protocol is required which unifies all features of the different protocol levels without compromising an efficient implementation. This work is dedicated to the design of the new Unified Layer Protocol (ULP) providing a unified communication scheme which allows communication among all processing units at different levels of an HPC system. Initially, the main features of general protocols are analyzed in detail. Further, properties used by modern protocols use are introduced and their function is explained. The two protocols that are deemed most relevant, Hyper-Transport (HT) and Peripheral Component Interconnect Express (PCIe), are analyzed in detail regarding to the previously specified aspects. The insight gained through this analysis is incorporated into the development of the ULP. During the development process, first the structure of the ULP is defined and various parameters are determined. Special attention is turned on the feasibility in hardware and the scalability for large systems. The following comparison with HT and PCIe shows that the newly developed ULP usually provides superior performance, even when the effective communication distance moves close to the processor. Further work is dedicated to the hardware development which first gave the inspiration for the development of the ULP. The insights gained during the development of the ULP were integrated into the hardware. The results show that the ULP fulfills the demands for a protocol used in the field of HPC. This is achieved for both, the processor-near communication, as well as for the communication among different nodes. With the ULP the need for time and energy-consuming protocol conversions is eliminated, while the feasibility in hardware is obtained