Fermilab Computing Division

CS Document 4602-v0

Study of a Fine Grained Threaded Framework Design

Document #:
CS-doc-4602-v0
Document type:
Presentation
Submitted by:
Christopher D. Jones
Updated by:
Christopher D. Jones
Document Created:
09 Jan 2012, 16:31
Contents Revised:
09 Jan 2012, 16:31
Metadata Revised:
09 Jan 2012, 16:31
Viewable by:
  • Public document
Modifiable by:

Quick Links:
Latest Version

Abstract:
Traditionally, HEP experiments exploit the multiple cores in a CPU by having each core process one event. However, future PC designs are expected to use CPUs which double the number of processing cores at the same rate as the cost of memory falls by a factor of two. This effectively means the amount of memory per processing core will remain constant. This is a major challenge for LHC processing frameworks since the LHC is expected to deliver more complex events (e.g. greater pileup events) in the coming years while the LHC experiment's frameworks are already memory constrained. Therefore in the not so distant future we may need to be able to efficiently use multiple cores to process one event. In this presentation we will discuss a design for an HEP processing framework which can allow very fine grained parallelization within one event as well as supporting processing multiple events simultaneously while minimizing the memory footprint of the job. The design is built around the libdispatch framework created by Apple Inc. (a port for Linux is available) whose central concept is the use of task queues. This design also accommodates the reality that not all code will be thread safe and therefore allows one to easily mark modules or sub parts of modules as being thread unsafe. In addition, the design efficiently handles the requirement that events in one run must all be processed before starting to process events from a different run. After explaining the design we will provide measurements from simulating different processing scenarios where the processing times used for the simulation are drawn from processing times measured from actual CMS event processing.
Files in Document:
None
Keywords:
CMS multicore
Associated with Events:
CHEP 2012 held on 21 May 2012 in New York, New York
DocDB Home ]  [ Search ] [ Authors ] [ Events ] [ Topics ]

DocDB Version 8.8.9, contact Document Database Administrators