Novel, highly-parallel software for the online storage system of the ATLAS experiment at CERN: Design and performances

The ATLAS experiment observes proton-proton collisions delivered by the LHC accelerator at CERN. The ATLAS Trigger and Data Acquisition (TDAQ) system selects interesting events on-line in a three-level trigger system in order to store them at a budgeted rate of several hundred Hz, for an average event size of ~1.5 MB. This paper focuses on the TDAQ data-logging system and in particular on the implementation and performance of a novel software design, reporting on the effort of exploiting the full power of multi-core hardware. In this respect, the main challenge presented by the data-logging workload is the conflict between the largely parallel nature of the event processing, including the recently introduced on-line event-compression, and the constraint of sequential file writing and checksum evaluation. This is further complicated by the necessity of operating in a fully data-driven mode, to cope with continuously evolving trigger and detector configurations. In this paper we will briefly discuss our development experience using recent concurrency-oriented libraries. We will then concentrate on the results of performance measurements performed on the current data-logging hardware. We will show that, even in the worst workload, the new parallel design is able to compete with the previous single-threaded one, while it is outperforming it in more favourable, realistic workloads. We will as well demonstrate the minimal overhead introduced by the above parallel techniques, considering the whole data-logging software performances with respect to the bare processing speed on the same hardware. Finally, we will discuss the effects of simultaneous multi-threading technologies, as found on recent CPUs. The data-logging operation in fact, mixing data processing and I/O, allows to efficiently exploit the features provided by these technologies.