solvers.OrchestrationModel

solvers.OrchestrationModel()

Analyzes Cluster Orchestration and Queueing (Little’s Law).

Caveat: This model uses a pedagogical M/D/1 queue (single server, deterministic service) to establish macroscopic wait-time bounds for dedicated, monolithic cluster workloads. For detailed multi-tenant job packing and preemption, a discrete-event M/G/c scheduler simulation would be required.

This model simulates the ‘Wait Wall’ in shared research clusters, calculating job completion times and researcher wait times based on cluster utilization and arrival rates.

Literature Source: 1. Little (1961), “A Proof for the Queuing Formula: L = λW.” 2. Barroso et al. (2018), “The Datacenter as a Computer” (Cluster Mgmt). 3. Jeon et al. (2019), “Analysis of Large-Scale Multi-Tenant GPU Clusters.”

Methods

Name Description
solve Solves for cluster wait times and utilization.

solve

solvers.OrchestrationModel.solve(
    fleet,
    arrival_rate_jobs_per_day,
    avg_job_duration_days,
)

Solves for cluster wait times and utilization.

Parameters

Name Type Description Default
fleet Fleet The hardware cluster configuration. required
arrival_rate_jobs_per_day float λ: Rate at which new training jobs are submitted. required
avg_job_duration_days float The average time a job takes to run if it has the whole cluster. required

Returns

Name Type Description
Dict[str, Any] Wait time, system length, and utilization metrics.
Back to top