Implementation of the ATLAS Trigger Within the ATLAS Multi-threaded Software Framework AthenaMT

We present an implementation of the ATLAS High Level Trigger (HLT) that provides parallel execution of trigger algorithms within the ATLAS multi-threaded software framework, AthenaMT. The original ATLAS software framework, Athena, was designed to process events serially and run algorithms sequentially on each event. Utilisation of many-core computer processors was achieved by creating multiple worker processes from a common parent process, each processing events sequentially. The current trend in computing hardware is to progressively increase the CPU core and hyper-thread count. If these trends continue as expected the number of worker threads and corresponding memory requirements will become unaffordable. To address this, the AthenaMT framework [1]- [3] is being developed as a common framework for both trigger and offline reconstruction. This framework enables parallelisation of algorithms within an event in addition to event-wise parallelisation. This offers the potential to decrease memory usage by reducing the number of events processed in parallel (events in-flight) to less than the number of threads. A further memory reduction can be achieved if algorithms can be made reentrant, such that a single algorithm instance can be executed concurrently in different threads, rather than having separate clones of the algorithms. The key features of the AthenaMT framework are that algorithms are scheduled automatically by the framework as directed by a precedence rules graph that is created from data dependencies between the algorithms, and by explicit Control Flow directives defined at run time. Algorithms execute in their own thread, from a shared thread pool. As many Algorithms are run simultaneously as there are available threads, and these Algorithms may execute in the context of the same or different events. Algorithms running concurrently in different threads must be reentrant (i.e. have no internal state) or clone-able so that instances of the same algorithm can