Indico has been migrated. Please contact support if you find anything missing.

22–23 Sept 2015
Europe/Copenhagen timezone
Starts
Ends
Europe/Copenhagen
Østervoldgade 5-7 DK-1350 Copenhagen

Background

Xeon-Phi many-core accelerators have been developed by Intel for massively parallel applications. A single Xeon-Phi processor contains 60 cores, which each can run 4 hardware threads, for a total of 240 threads. Each core has a simple in-order architecture, but also a very wide 512-bit vector unit resulting in a high peak floating-point operations capability. The Xeon-Phi is in principle much more accesible than a GPU. Each Xeon-Phi card looks like a very large SMP node with 240 logical processors, 8GB of memory, and a barebones Linux interface. It is easy to port applications to Xeon-Phis, but reaching something close to peak performance requires careful selection of parallel algorithms, vectorized code, and reuse of cached data. At the University of Copenhagen HPC center the theoretical and computational astrophysics groups recently installed a large new cluster, which contains a sub-section with 39 accelerated nodes each equipped with two 60-core Xeon-Phi cards and 20 ivy-bridge cores, and a fast FDR infiniband network.
 

Topics

  • Xeon-Phi basics and architecture
  • Xeon-Phi vectorization and threading
  • Xeon-Phi offloading (OpenMP 4+, pyMIC)
  • Xeon-Phi tools (focus on VTune)
  • Hands-on exercises
  • Local contributions
 

Objective

At the Xeon-Phi days we will have lectures by Intel expert Mikko Byckling about the Xeon-Phi architecture, hands-on exercise sessions, and locally contributed talks. It will give participants a general idea of how Xeon-Phi's work, and an understanding of how likely it is their own programs and algorithms can be efficiently executed on the Xeon-Phi.

Pre-requisites

During the Xeon-Phi days a special queue will be created, which gives all participants access to the Xeon-Phi cluster to carry out the exercises. This requires an account at the HPC center. The second day we will use a graphical performance tool (Intel Vtune) that is best run through a vnc session, when used on a remote computer. We recommend that you have a VNC client installed.

You are welcome to try your own application on the Xeon-Phis. This can be done in two different manners:
  • Native mode: By ssh'ing to the Xeon-Phi card you can execute the code directly on the accelerator. Remember that the environment is peculiar, with access to 240 cores, but only 8GB of memory. To run on the nodes requires a recompilation of the code with the Intel compilers using a specific flag "-mmic", but otherwise most codes, that do not rely on extensive third party, should be easy to get running.
  • Offload mode: You instrument the code with specific pragmas / directives, specifying "off-loaded" regions, which will be executed on the Phi, while the rest runs on the host node. This requires changes to the source code, but can be easier if e.g. the code relies on third party libraries.
Both modes of execution will be discussed during the workshop.


Registration and contributions

The Xeon-Phi days will be held at the Centre for Star and Planet Formation. NOTICE there has been a change of lecture room! The lecture room is on the third floor of the DNA Lab on Øster Farimagsgade 2D. Follow the signs, and go in through entrance E, take the elevator to third floor, and turn right. The entrance is at the end of the corridor. The lunch will be served in the canteen at StarPlan.



Local talks are very welcome. Contact the organizers if you would like to contribute to the meeting.

Background Material

We suggest that you take a look at the following two documents beforehand, to familiarize yourself with basic aspect of the Xeon-Phi accelerator, and get the most out of the meeting

Intel Xeon-Phi Coprocessor Developer's Quick Start Guide
 
https://software.intel.com/en-us/articles/intel-xeon-phi-coprocessor-developers-quick-start-guide
 
PRACE Xeon-Phi best practice guide
 
http://www.prace-ri.eu/best-practice-guide-intel-xeon-phi-html/
 

Organizers

  • Troels Haugbølle
  • Jon Ramsey
  • Åke Nordlund
  • Martin Pessah
  • Brian Vinter
The Xeon-Phi days are supported by