version française rss feed
HAL : in2p3-00700283, version 1

Fiche concise  Récupérer au format
I/O Strategies for Multicore Processing in ATLAS
van Gemmeren P., Binet S., Calafiura P., Lavrijsen W., Malon D. et al
Dans Journal of Physics: Conference Series - Computing in High Energy and Nuclear Physics (CHEP2012), New York : États-Unis (2012) - http://hal.in2p3.fr/in2p3-00700283
Informatique/Architectures Matérielles
I/O Strategies for Multicore Processing in ATLAS
P. van Gemmeren1, S. Binet2, P. Calafiura3, W. Lavrijsen3, D. Malon1, V. Tsulaia3
1 :  ANL - Argonne National Laboratory
US Department of Energy – University of Chicago
9700 S. Cass Avenue Argonne, IL 60439
2 :  LAL - Laboratoire de l'Accélérateur Linéaire
CNRS : UMR8607 – IN2P3 – Université Paris XI - Paris Sud
Centre Scientifique d'Orsay B.P. 34 91898 ORSAY Cedex
3 :  LBNL - Lawrence Berkeley National Laboratory
Lawrence Berkeley National Lab
Lawrence Berkeley National Lab MS 50F Cyclotron Rd., Berkeley, CA 94720
A critical component of any multicore/manycore application architecture is the handling of input and output. Even in the simplest of models, design decisions interact both in obvious and in subtle ways with persistence strategies. When multiple workers handle I/O independently using distinct instances of a serial I/O framework, for example, it may happen that because of the way data from consecutive events are compressed together, there may be serious inefficiencies, with workers redundantly reading the same buffers, or multiple instances thereof. With shared reader strategies, caching and buffer management by the persistence infrastructure and by the control framework may have decisive performance implications for a variety of design choices. Providing the next event may seem straightforward when all event data are contiguously stored in a block, but there may be performance penalties to such strategies when only a subset of a given event's data are needed; conversely, when event data are partitioned by type in persistent storage, providing the next event becomes more complicated, requiring marshaling of data from many I/O buffers. Output strategies pose similarly subtle problems, with complications that may lead to significant serialization and the possibility of serial bottlenecks, either during writing or in post-processing, e.g., during data stream merging. In this paper we describe the I/O components of AthenaMP, the multicore implementation of the ATLAS control framework, and the considerations that have led to the current design, with attention to how these I/O components interact with ATLAS persistent data organization and infrastructure.

Communications avec actes
Journal of Physics: Conference Series
Institute of Physics Publishing

Computing in High Energy and Nuclear Physics (CHEP2012)
New York
P. Van Gemmeren

LAL 12-179