ASTRA-sim is a distributed AI system simulator. It models the end-to-end software and hardware stack of modern AI systems - encompassing workload scheduling, collective communication algorithms, and hardware architectures (compute/memory/network). Through a suite of APIs, it enables plug-and-play of external open/proprietary components for modeling different parts of the AI system. This provides end-to-end multi-fidelity simulation capabilities for aiding in design and deployment of next-generation distributed AI systems.
This repository includes a unified analytical layer for three inference studies:
tp_pp_crossoverSweeps TP versus PP tradeoffs for 70B-class dense models.serving_disagg_colocatedRuns the existing disaggregated-versus-colocated serving simulator through the analytical frontend without changing TTFT, TPOT, E2E, queueing, or goodput behavior.attention_ffn_disaggregationCompares homogeneous GPU against heterogeneous GPU+LPU placement for trillion-parameter MoE decode.
The shared implementation lives under astra-sim/analytical and keeps model specs, hardware specs, interconnects, cost models, and result writers in one framework-native module.
Build the analytical binaries with:
./build/astra_analytical/build.shThe main binaries are:
build/astra_analytical/build/bin/AstraSim_Analytical_Congestion_Awarebuild/astra_analytical/build/bin/AstraSim_Analytical_Congestion_Unaware
Run the shipped analytical configs with:
./build/astra_analytical/build/bin/AstraSim_Analytical_Congestion_Aware \
--analytical-config=$(realpath configs/tp_pp_70b.yaml)
./build/astra_analytical/build/bin/AstraSim_Analytical_Congestion_Aware \
--analytical-config=$(realpath configs/serving_disagg_colocated.yaml)
./build/astra_analytical/build/bin/AstraSim_Analytical_Congestion_Aware \
--analytical-config=$(realpath configs/attention_ffn_gpu_lpu_moe.yaml)Important behavior:
--analytical-configis the public entrypoint for these analytical modes.tp_pp_crossoverandattention_ffn_disaggregationare closed-form analytical runs and do not instantiate the event-driven serving runtime.serving_disagg_colocatedstill uses the existing serving simulator underneath, so serving behavior stays on the established path.
The checked-in examples live in configs:
These YAMLs are fully explicit. Output paths are resolved relative to the YAML file location, which makes it easy to copy a config and redirect outputs without rewriting absolute paths.
tp_pp_crossover writes a row-oriented CSV plus a summary JSON containing pure-TP versus pure-PP crossover records.
serving_disagg_colocated keeps the existing request metrics CSV, summary JSON, and metadata JSON format from the serving simulator.
attention_ffn_disaggregation writes a row-oriented CSV plus a summary JSON comparing:
homogeneous_gpuheterogeneous_gpu_lpu
Run the full regression suite with:
./tests/run_all.shRun the analytical regression directly with:
./tests/rt_analytical/run.shRun the analytical logic tests directly with:
./build/astra_analytical/build/bin/AstraSim_Analytical_Logic_TestsFor a focused usage guide covering config structure, commands, outputs, and extension points, see docs/project/analytical-guide.md.
Here is a concise visual summary of ASTRA-sim, showing its layers and APIs:
For a comprehensive understanding of the tool, and to gain insights into its capabilities, please visit our website.
For information on how to use ASTRA-sim, please visit our Wiki.
ASTRA-sim accepts MLCommons Chakra Execution Traces as workload-layer inputs. For details, please visit Chakra Github.
ASTRA-sim is currently at version 2.0.
The previous version, ASTRA-sim 1.0, is available in the ASTRA-sim-1.0 branch.
We encourage community contributions to ASTRA-sim via PRs.
For any questions about using ASTRA-sim, you can email the ASTRA-sim User Mailing List: astrasim-users@googlegroups.com
To join the mailing list, please fill out the following form: https://forms.gle/18KVS99SG3k9CGXm6
We appreciate your interest and support in ASTRA-sim!