4.4 <strong>Scientific</strong> Computing, In<strong>for</strong>mation Technology 114
The ALICE High Level Trigger Collaborators: T. Kollegger 1 , T. Alt 1 , S. Gorbunov 1 , S. Kalcher 1 , M. Kretz 1 , M. Langhammer 1 , V. Lindenstruth 1 , D. Ram 1 , D. Rohr 1 , O. Smørholm 1 , T. Steinbeck 1 1 <strong>Frankfurt</strong> <strong>Institute</strong> <strong>for</strong> <strong>Advanced</strong> <strong>Studies</strong> ALICE is one of the four main experiments at the Large Hadron Collider (LHC) at the European Center <strong>for</strong> Particle Physics CERN, Geneva. Its main goal is to study the properties of the hot and dense medium created in collisions of heavy ions. In the autumn of <strong>2011</strong>, LHC was operating with lead (Pb) ions and achieved luminosities an order of magnitude higher than in previous years, with up to 6kHz of Pb+Pb collisions. This new record interaction rates posed a challenge <strong>for</strong> the read-out of the detector. Each Pb+Pb collision creates up to 80MB of total data volume from the various sub-detectors of ALICE, with the Time Projection Chamber (TPC) by far the largest (∼60MB). This data rate by far exceeds the per<strong>for</strong>mance of the detector read-out systems, so events of interest have to be selected, the task of the trigger system. Another bottleneck is the mass-storage system, whose sustained bandwidth is significantly lower than the maximum detector read-out bandwidth. The ALICE High Level Trigger (HLT) was designed to close the gap between the maximum detector read-out bandwidth and the storage bandwidth. Being the first point in ALICE where the data from all sub-detectors is available, the HLT reconstructs the full event, i.e. the trajectories of the several thousand particles emitted from the collission and passing through the detector system. Figure 1 shows a visualization of a fully reconstructed event. Figure 1: A central Pb+Pb event as reconstructed by the ALICE High Level Trigger. After sucessful operation in 2009 and 2010, the HLT was upgraded in <strong>2011</strong> to deal with the higher data rates. The particle track reconstruction in the HLT uses graphic cards (GPUs) in addition to the CPUs, the number of nodes with GPUs was doubled to a total of 64. The output bandwidth of the system was increased by a factor 4, now reaching more than 3GB/s. With these upgrades, the HLT was able to deal with the maximum possible detector read-out rates of more than 20 GB/s. A special emphasis in <strong>2011</strong> was the improvment of the physics reconstruction per<strong>for</strong>mance. For this the first processing step of the TPC reconstruction, the identification of charge clusters deposited by charged particles traversing the TPC gas, has been optimized. After optimization of the FPGA based HLT cluster finder algorithm and its parameters, detailed studies showed only negligibale differences in the final track reconstruction parameters compared to using the, by far slower, offline cluster finder. Having demonstrated this excellent physics per<strong>for</strong>mance, ALICE decided to discard the TPC raw data, relying <strong>for</strong> the particle reconstruction soley on the HLT cluster finder output. Together with data <strong>for</strong>mat optimizations, entropy-reducing trans<strong>for</strong>mations of the cluster data and subsequent lossless Huffman-compression, the HLT reduced the 20 GB/s input rate to 4 GB/s storage rate, enabling the recording of all events which could be handeld by the detector read-out system and thus maximizing the physics reach of ALICE. 115