DisCostiC v2.0.0
Distributed Cost in Cluster: a cross-architecture simulator for massively parallel MPI programs
Loading...
Searching...
No Matches
💡 Description

Part of the main README.

DisCostiC is a cross-architecture, resource-based parallel simulation framework for predicting the performance of MPI programs on current and future HPC systems.✨

It schedules computation and communication from a C++ domain-specific embedded language (DSEL) and evaluates them with analytical models: Roofline and ECM for computation, and LogGP and Hockney for communication. Machine descriptions are YAML files. Network descriptions are numeric .cfg tables (latency, inverse bandwidth, eager limit, and LogGP parameters). The simulation does not run the full application on the target hardware.⭐

The same DSEL covers MPI collective modeling, multi-kernel workflows, and optional heterogeneous clusters. Those placements are documented in Modes.🌟

The application blueprint uses concepts from modern C++ so domain experts can encode inter-process dependencies without taking on unknown effects from a particular machine.💥

⚙️ Framework workflow 💬

- Application model
* Domain knowledge expressed in the DisCostiC DSEL: call dependencies, computation (data access, flop count), and communication (volume, mode, and protocol), including point-to-point, collectives, multi-kernel phases, and heterogeneous placements
- Analytic first-principle performance models
* Computation models: Roofline and refined ECM models, including the memory-bandwidth bottleneck at chip level
* Communication models: LogGP and Hockney (latency-bandwidth) for intra-chip, inter-chip, and inter-node scopes
- System models
* YAML machine files and numeric network `.cfg` tables for the full topology: cores, chips, nodes, network interfaces, and clusters
+ DisCostiC combines all of them to simulate the runtime cost of distributed applications without executing the full application on the target hardware.

🖍 Advantages over existing tools 💬

- DSEL approach for creating an application blueprint
* Inter-process dependencies, collectives, multi-kernel phases, and heterogeneous clusters, without introducing unknown consequences from the actual systems
- A full-scale first-principle-model-based simulator
* Roofline/ECM and LogGP/Hockney across cores, chips, nodes, and networks
- Efficient speed
* No intermediate application traces, unlike offline trace-driven tools
* No requirement to execute the application on the host architecture
+ An open-source, low-entry simulator for model-based design-space exploration