Heidelberg 2015 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

HK: Fachverband Physik der Hadronen und Kerne

HK 69: Instrumentation 19

HK 69.7: Talk

Friday, March 27, 2015, 16:15–16:30, M/HS1

Studies on high performance Timeslice building on the CBM FLES — •Helvi Hartmann and V. Lindenstruth for the CBM collaboration — Frankfurt Institute for Advanced Studies, Goethe University, Frankfurt, Germany

In contrast to already existing high energy physics experiments the Compressed Baryonic Matter (CBM) experiment collects all data untriggered. The First-level Event Selector (FLES), which denotes a high performance computer cluster, processes the very high incoming data rate of 1 TByte/s and performs a full online event reconstruction. For this task it needs to access the raw detector data in time intervals referred to as Timeslices. In order to construct the Timeslices, the FLES Timeslice building has to combine data from all input links and distribute them via a high-performance network to the compute nodes.

For fast data transfer the Infiniband network has proven to be appropriate. One option to address the network is using Infiniband (RDMA) Verbs directly and potentially making best use of Infiniband. However, it is a very low-level implementation relying on the hardware and neglecting other possible network technologies in the future. Another approach is to apply a high-level API like MPI which is independent of the underlying hardware and suitable for less error prone software development.

I would like to present the given possibilities and to show the results of benchmarks ran on high-performance computing clusters. The solutions are evaluated regarding the Timeslice building in CBM.