The DEEP/-ER platform is more flexible than a standard architecture: It enables different usage models:

  • Dynamic ratio of processors/coprocessors
    • Each process in the Cluster Nodes has its own set of coprocessors in the Booster Nodes (determined at runtime).
    • Applications that rely mostly on accelerators do not have the host processors idling a significant amount of time.
  • Use Booster as pool of accelerators (globally shared)
    • All the processes in the Cluster Nodes have access to all the coprocessors in the Booster Nodes.
    • This helps to alleviate load imbalance.
  • Discrete use of the Booster
    • Autonomous use of Xeon Phis
    • e. g. interesting for MC and QCD applications
  • Discrete use + I/O offload
    • Discrete use of the Booster Nodes
    • Use Cluster as I/O proxy
  • Specialized symmetric mode
    • Mix of Cluster and Booster Nodes without “hierarchical” relationship, but each part of the system runs a specific part of the application.
    • The code division enables new possibilities e. g. reduced noise by using large number of particles.

Favorable side effect: These usage models enable a more efficient use of system resources!

  • Don’t bother about heterogeneous runtime characteristics: The DEEP/-ER systems are especially suited for HPC applications with code parts that differ in their scalability characteristics (i.e. multi-scale and/or multi-physics simulations).
  • Ease of use: Porting an application is a snap with our programming model – especially if you are using MPI already. The model is based on standards, protecting your investment in code modernization. Check out the guidelines here.
  • Fewer failures: The system software provides advanced features for resiliency, reducing the overhead of task and process checkpoints by exploiting a multi-level memory hierarchy.