IBM ExaBounds models the performance and power of a large-scale high-performance computer (HPC). It is an analytical model (as opposed to a simulator), i.e., the interactions between system components are described mathematically rather than functionally. IBM ExaBounds comprises detailed performance models of multi-core processors, memory hierarchies and interconnect networks as well as the IBM Memory-Scheduler-Agnostic Power Model for DRAMs (also stand-alone). To model processor power it invokes McPAT.
The input to the system model are a set of JSON files describing the hardware configuration and a software profile. Software profiles can be generated by the IBM Platform-Independent Software Analysis tool and extrapolated to the target scale with the IBM Exascale Extrapolator. (The latter is included in this IBM ExaBounds repository.) The IBM Appresentor automates the design of experiments and profiling required for extrapolation.
The complete tool flow is this:
- IBM Appresentor, kicking off IBM Platform-Independent Software Analysis
- IBM Exascale Extrapolator
- IBM ExaBounds
Setting up and using the tools is described in the documentation for IBM ExaBounds and the IBM Exascale Extrapolator.
This tool chain was developed by the Algorithms & Machines team at IBM Research – Zurich in the context of Dome, a joint program between IBM Research – Zurich and Astron, the Netherlands Institute for Radio Astronomy.
The following main publications present the science behind IBM ExaBounds and the IBM Exascale Extrapolator:
- Rik Jongerius, Andreea Anghel, Gero Dittmann, Giovanni Mariani, Erik Vermij, Henk Corporaal: "Analytic multi-core processor model for fast design-space exploration." IEEE Transactions on Computers, 2018.
- Andreea Anghel: "On Large-Scale System Performance Analysis and Software Characterization." Ph.D. Thesis, ETH Zurich, 2017. (For interconnect and branch-prediction modeling, full-system accuracy analysis. Conference publications in preparation.)
- Sandeep Poddar, Rik Jongerius, Leandro Fiorin, Giovanni Mariani, Gero Dittmann, Andreea Anghel, and Henk Corporaal: "MeSAP: A fast analytic power model for DRAM memories." IEEE/ACM Design Automation and Test in Europe (DATE), 2017, pp. 49-54.
- Giovanni Mariani, Andreea Anghel, Rik Jongerius, Gero Dittmann: "Scaling Properties of Parallel Applications to Exascale." International Journal of Parallel Programming 44(5), 975--1002, Springer, 2016.
- Giovanni Mariani, Andreea Anghel, Rik Jongerius, Gero Dittmann: "Classification of Thread Profiles for Scaling Application Behavior." Parallel Computing, Elsevier, 2017.