A comparison of CMS Tier0-dataflow scenarios using the Yasper simulation tool

The CMS experiment at CERN will produce large amounts of data in short time periods. Because the data buffers at the experiment are not large enough, this data needs to be transferred to other storages. The CMS Tier0 will be an enormous job processing and storage facility at the CERN site. One part of this Tier0, called the Tier0 input buffer, has the task to readout the experiment data buffers and to supply these data to other tasks that need to be carried out with it (such as storing). It has to make sure that no data is lost. This thesis compares different scenarios to work with a set of disk servers in order to accomplish the Tier0 input buffer tasks. To increase the performance per disk server, write and read actions on the same disk server are separated. To find the optimal moments a disk server should change from accepting and writing items to supplying items to other tasks, the combination of various parameters, such as the usage of a particular queuing discipline (like FIFO, LIFO, LPTF and SPTF) and the state of the disk server has been studied. To make the actual comparisons a simulation of dataflow models of the different scenarios has been used. These simulations have been performed with the Yasper simulation tool. This tool uses Petri Net models as its input. To be more certain that the models represent the real situation, some model parts have been remodelled in a tool called GPSS. This tool is not using Petri Nets as its input model; instead it uses queuing models described in a special GPSS language. The results of the simulations show that the best queuing discipline to be used with the Tier0 input buffer is the LPTF discipline. In particular in combination with a change moment as soon as a disk server has been readout completely.