Turing : Introduction

Turing is the massively parallel architecture of IDRIS. This is an IBM Blue Gene/Q machine which is particularly well-balanced (network, I/O, computing power and memory bandwidth) and has a remarkable energy efficiency (2.17 Gflops/W).

Hardware characteristics of Turing :

  • 6 racks, each containing 1,024 compute nodes with 16 cores per node, for a total of 98,304 cores
  • Cumulated peak performance of 1.258 Pflop/s (42nd in the TOP500 worldwide ranking of November 2014)
  • 96 TiB of total memory (1 GiB per execution core)
  • GPFS parallel file system (WORKDIR) shared by Ada, Adapp and Turing, with 50 GiB/s bandwidth in write and in read.

The codes which run on Turing must have a sufficiently high degree of parallelism (ideally, at least 512 or more execution cores) and not go beyond the 1 GiB memory constraint per core. In production, there can be up to 4 racks of accessible resources (or 65,536 execution cores). Specific classes (of 1 to 64 compute nodes) are available for the development or code porting phases. You may consult the batch limits of the class structures for Turing.

For more information concerning the available software or librairies, the architecture, and the usage of the Turing machine resources, see here.