MARLIN Simplified Nash Equilibrium PoC

What We Tested

The MARLIN paper (arXiv:2605.13496) proposes a multi-agent game-theoretic RL framework where three agents (optimizing TTFT, carbon, and water/cost) reach a Nash equilibrium over LLM inference scheduling across geo-distributed cloud datacenters.

Effloow Lab built a simplified Python stdlib reproduction of the core trade-off logic:

• 3 agents, each with a utility function aligned to one objective (speed, carbon, cost)
• 3 simulated datacenter regions with distinct carbon intensity, water usage, cost, and base TTFT profiles
• Brute-force Nash equilibrium search over discrete 0.1-step allocation fractions

This is NOT a reproduction of MARLIN's actual RL training. It models the strategic structure (Pareto trade-offs, Nash stability) without any learned policies.

Commands Run

python3 --version
# Python 3.12.8

python3 /tmp/marlin_nash_poc.py

Script written to /tmp/marlin_nash_poc.py. Uses only: math, itertools.

Region profiles used:

• us-east-1: 420 gCO2/kWh, 1.8 L/kWh water, $0.10/kWh, 120ms base TTFT
• eu-west-2: 85 gCO2/kWh, 0.4 L/kWh water, $0.18/kWh, 145ms base TTFT
• ap-south-1: 680 gCO2/kWh, 2.3 L/kWh water, $0.07/kWh, 110ms base TTFT

Output

============================================================
Effloow Lab — MARLIN Nash Equilibrium PoC
3 agents, 3 datacenter regions, stdlib only
============================================================

[NAIVE] Route all to ap-south-1 (cheapest)
  Allocation : {'us-east-1': 0.0, 'eu-west-2': 0.0, 'ap-south-1': 1.0}
  TTFT       : 66.0 ms
  Carbon     : 680.0 gCO2/kWh
  Water      : 2.30 L/kWh
  Cost       : $0.070/kWh

[NASH EQUILIBRIUM] Stable multi-agent allocation
  Allocation : {'us-east-1': 0.0, 'eu-west-2': 1.0, 'ap-south-1': 0.0}
  TTFT       : 87.0 ms
  Carbon     : 85.0 gCO2/kWh
  Water      : 0.40 L/kWh
  Cost       : $0.180/kWh

[IMPROVEMENT vs NAIVE]
  TTFT reduction    : -31.8%
  Carbon reduction  : 87.5%
  Water reduction   : 82.6%

[NOTE] Paper claims vs naive baselines (full ML training):
  TTFT >= 18% | Carbon >= 33% | Water >= 43%
  This PoC models the TRADE-OFF LOGIC, not the trained RL policy.

============================================================

No errors. Exit code 0.

What Worked

Successfully demonstrated the Pareto trade-off geometry at the core of MARLIN's argument:

1. Nash stability confirmed: The equilibrium search found eu-west-2 as the stable single-region allocation — no agent can unilaterally improve by deviating to another region without another agent's utility declining.

2. Carbon/water reduction confirmed: The 8x carbon intensity gap between ap-south-1 (680 gCO2/kWh) and eu-west-2 (85 gCO2/kWh) drives 87.5% carbon and 82.6% water reductions vs naive cost-optimized routing.

3. Structural tension visible: The Nash equilibrium routes 100% to eu-west-2 at the cost of TTFT (87ms vs 66ms). This illustrates the exact problem MARLIN's RL training solves: a greedy single-region Nash collapses to one objective; trained agents find multi-region mixed allocations that improve TTFT and carbon simultaneously.

Limitations

1. Single-region collapse: Our simplified model routes 100% to eu-west-2. MARLIN's trained agents spread load across regions to simultaneously improve TTFT while reducing carbon — something binary Nash cannot do.

2. No RL training: MARLIN uses multi-agent RL (policy gradients) trained over thousands of datacenter simulation steps. This PoC uses static brute-force search with hand-coded utility functions.

3. Simplified TTFT model: We model TTFT as a linear function of allocation fraction. Real TTFT depends on queue depth, KV cache hit rates, prefill batch size, network latency, and GPU utilization curves.

4. No temporal dynamics: Real datacenter carbon intensity shifts by hour (solar availability, grid mix). MARLIN's agents adapt continuously; ours is a static snapshot.

5. 3 regions vs production scale: Production LLM inference infrastructure spans 20+ availability zones with complex cross-region dependencies.

6. No water cooling model: The paper claims 43% water reduction. Our PoC models water as a fixed per-kWh rate per region, not a dynamic cooling model tied to ambient temperature and datacenter load.

Verdict

The core Pareto trade-off claim is highly plausible. Carbon intensity varies 8x across realistic grid regions — this is documented fact, not a modeling assumption. The simplified PoC confirms that even without RL training, carbon-aware routing produces substantial sustainability improvements. The 33% carbon and 43% water reduction claims in the paper are conservative relative to what pure carbon-routing achieves in our model; the technically interesting part is MARLIN's simultaneous 18% TTFT improvement, which requires RL-trained agents to find mixed-strategy load distributions that a greedy heuristic misses. Full validation would require running MARLIN's actual RL training pipeline against a datacenter simulator with real grid carbon data and live LLM inference traffic.