Repository navigation
With great pleasure, version 2.0.0 of DisCostiC has been released.
It is a cross-architecture resource-based parallel simulation framework that can efficiently predict the performance of real or hypothetical massively parallel MPI programs on current and future heterogeneous systems. This release extends the simulator beyond point-to-point communication to MPI collectives, multi-kernel and heterogeneous workflows, and validates its predictions against measurements.
New features
- Collective mode with analytic models for broadcast (binomial, ring), reduce (binomial) and allreduce (recursive doubling)
- Multi-kernel mode for chaining two compute phases per timestep
- Heterogeneous mode for simulating ranks across two different systems
- New machine models for Intel Sapphire Rapids and AMD Zen5 EPYC 9965
- New network models for Sapphire Rapids and updated network models for Ice Lake, Broadwell and A64FX
- Automated machine-file generation with likwid-bench via SLURM
- Trace output of the simulated application in ITAC (STF) format, in addition to the existing Chrome (JSON) and Vampir (OTF2) formats
- New benchmarks: STREAM collective, Pi solver + STREAM, ray tracing + optical flow, and two-kernel workflows
- Doc: New Doxygen site, README split into topic guides under doc/, plus an updated wiki page
- Added
demowith five self-contained case studies: STREAM triad, 2D Jacobi heat, Pi solver + STREAM, ray tracing + optical flow, and MPI collectives. The four application demos cover ECM modeling, single and multi-timestep runs, delay propagation, and strong/weak scaling up to 32 nodes; the collectives demo compares simulated collective costs with IMB measurements