############################################################
                    Seminar Announcement

            School of Computer Science, CECS
             The Australian National University
############################################################

Date:  Thursday, 19 Feb 2009
Time:  4:00 pm to 5:00 pm
Venue: Room N101, CSIT Building [108]


Speaker: Danny Robson

Title: Efficient and Accurate Performance Evaluation of Multi-core NUMA Clusters
	(Thesis Proposal Review seminar)


Abstract:

  In recent years there has been increasing investment in large,
  parallel, compute clusters; particularly those with more capable,
  parallel processor technology. These systems have proved useful in
  executing highly numerical workloads in various scientific fields.
  
  As hardware becomes more complex, it becomes increasingly difficult to
  accurately predict system performance and discover attributes that
  contribute significantly to workload execution. This is especially
  apparent when utilising Chip Multi-Processing (CMP) and Non-Uniform
  Memory Access (NUMA) architectures. Due to this complexity, optimal use
  of these systems requires use of performance analysis tools, on both
  individual nodes and clusters as a whole.
  
  Existing analysis tools employ a variety of techniques with a range of
  efficiency, accuracy and precision characteristics; analytical methods,
  software and hardware profiling, and emulation and simulation are all
  widely applied in many contexts. Dynamic Binary Translation (DBT) and
  Dynamic Binary Instrumentation (DBI) techniques allow an efficient
  means of adding instrumentation code to arbitrary executables at
  runtime. These allow a tool developer to insert arbitrary code which
  drives performance analysis models during guest execution.
  
  By investigating simulation techniques for workload and system
  analysis, we will create an accessible means for performance
  evaluation. The focus will be on assisting developer analysis of highly
  parallel, numerical applications through precise and easily accessible
  feedback on their application's execution. This talk focuses on our
  existing and proposed performance analysis methods.
  
  We will discuss the tools we are developing with the Valgrind DBI
  framework, including modifications which allow parallel execution of
  guest threads with associated execution speedup. To enable efficient
  analysis of single cluster node we are constructing an improved cache
  simulator and NUMA emulation layer within our analysis tool, NUMAgrind.
  
  We will outline future work in the addition of timing accuracy to the
  NUMAgrind tool, for single nodes and full clusters. In the single node
  case we outline methods for coupling existing processor models and
  simulators to the memory system for timing estimation. In the cluster
  case, we describe coupling timing estimators for network communications
  and how the results will moderate the timing of guest execution.

Biography:

  Danny received a Bachelor of Engineering (Software) from the University
  of New South Wales in 2007. He is currently a PhD student at the
  School of Computer Science associated with the CC-NUMA Project.


URL: http://cs.anu.edu.au/lib/seminars/seminars09/dept20090219




############################################################
Seminars homepage:  http://cs.anu.edu.au/seminars/

If you like to give a seminar please contact:
                           seminars-owner [at] cs.anu.edu.au
############################################################