Research And Implementation Of InfiniBand Baseboard Management

As an interconnection technology, InfiniBand is considered as a new solution to eliminate the performance bottleneck of current I/O framework because it owns many advantages of high bandwidth, low delay, etc. InfiniBand subnet implements high speed interconnection and data transmission from tens of nodes to hundreds of nodes. Now, InfiniBand technology is widely used in high performance computing and becomes the first choice for the interconnection in high performance computing.Baseboard Management is one important component in the InfiniBand network management because its main function is to monitor and manage the hardware in the InfiniBand subnet devices to ensure system stability.Three main contributions of this thesis are listed as follows:(1) Study the infrastructure of InfiniBand, analyze and discuss in depth about the management mechanism of InfiniBand, particularly, the Baseboard Management in General Services Management;(2) Based on the analysis of the management mechanisms of IBS216Q switch, we propose and implement a management model in the I2C network switch. Moreover, we test the board temperature in the switch based on this model. The efficiency of this system is also verified;(3) Study the routing algorithm used in the InfiniBand subnet with Fat-tree topology and analyze the modes of failures. We propose a metric to measure the impact scale of the subnet failure. Based on this metric, we analyze the failure mode of m-port 3-trees InfiniBand subnet and present a solution, which minimizes the bad effect of the subnet failure. At last, we take a step further in the analysis and study of the failure mode and impact scale to subnet from switches;In a word, this thesis studies in-depth, designs and implements the mechanisms of InfiniBand Baseboard Management. Meanwhile, we analyze the failure mode of the InfiniBand subnet with fat-tree topology. We believe that our work will be very useful to improve the failure-tolerance ability of InfiniBand network and build high performance cluster systems based on InfiniBand technology.