Interconnect-Aware Topology Simulator for Model Serving with Google ADK
Simulate network interconnect bottlenecks in multi-node TPU and GPU clusters for large model inference. Construct a multi-agent cluster topology simulator in Google ADK that models NVLink vs PCIe interconnect latency under tensor-parallel and pipeline-parallel workload distributions.
What you are building
Build an interconnect simulation pipeline in Google ADK to quantify cross-node tensor parallel communication overhead across different hardware topologies.
How work is evaluated
Evaluates calculated communication overhead latency against mathematical baseline models.
Datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
Math Precision Test
Verify calculation accuracy within tolerance.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Simulation Accuracy Rate
Alignment with analytical networking benchmarks. • target: 0.98 • range: 0-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Learning goals
Build interconnect topology models in Google ADK for TPU v5e/v6 and NVIDIA HGX clusters.
Simulate AllReduce and AllGather collective communication delays under multi-node splitting.
Recommend optimal TP/PP layout strategies based on physical link bandwidths.
Resources
Network topology map dataset for standard 8-GPU HGX nodes and TPU pod slices.
Challenge setup
Evaluates calculated communication overhead latency against mathematical baseline models.
Challenge input
JSON tensor_size_mb, interconnect_type (nvlink|pcie|ethernet), bandwidth_gbps
Google ADK
Google Agent Development Kit framework.
Evaluated output
JSON estimated_comm_time_ms
- Verify calculation accuracy within tolerance.
- Alignment with analytical networking benchmarks. • target: 0.98 • range: 0-1
- Benchmark: Distributed Parallelism Communication Benchmark
- Simulation Accuracy Rate target: 98%
- 1 public reference case
- Python execution harness
- Python sandbox (unavailable on Versalist)
Shows the first task, up to three tools, and up to three evidence items.
View technical recipe
Configured tools
- Google · Required
- Google ADK · Optional
- Google Jules · Optional
- Google Jules · Optional
Evaluation contract
- Math Precision Test · Weight 1
- Simulation Accuracy Rate · Weight 1
Metrics
- Simulation Accuracy Rate · Target 0.98
Benchmarks
- Distributed Parallelism Communication Benchmark
- general
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: JSON tensor_size_mb, interconnect_type (nvlink|pcie|ethernet), bandwidth_gbps
Output: JSON estimated_comm_time_ms
Execution harness: Python
Evaluation tools
- Google ADK · Agent Framework
Recipe freeze marker
No recipe freeze time is recorded.