Mcds Scheduling Eval

mcds-scheduling-eval

MCDS: AI Augmented Workflow Scheduling in Mobile Edge Cloud Computing Systems — Tuli et al. (2021) (arXiv:2112.07269, 2021)

What this evaluates

Evaluates AI-augmented workflow scheduling in mobile edge-cloud environments by comparing execution time, energy consumption, SLA violations, and fairness against state-of-the-art baselines under dynamic workloads and host mobility.

Datasets

• WFCommons (Pegasus workflows: BLAST, Cycles, Montage) — total ?; splits: benchmark (-1)

Metrics

• SLA Violation Rate (primary) — range: percent
  • Fraction of workflows whose response time exceeds the application-specific deadline. The deadline is defined as the 98th percentile response time of the Closure baseline for that application.
• Average Response Time — range: other
  • Mean time from workflow submission to completion across all executed workflows.
• Average Energy Consumption — range: other
  • Normalized average power usage of the edge-cloud environment per scheduling interval.
• Scheduling Fairness — range: [0, 1]
  • Jain’s fairness index computed over the Instructions Per Second (IPS) of running tasks.

Input / output format

Input: Workflow DAG specifications (tasks, precedence constraints, resource requirements) and real-time host resource utilization metrics (CPU, RAM, disk, bandwidth) at each scheduling interval.

Output: Task-to-host assignment mapping (placement decisions) for each workflow in the scheduling interval, including migration decisions if applicable.

Scoring recipe

def compute_metrics(predictions, gold, baseline_closure):
    sla_deadline = np.percentile(baseline_closure['response_time'], 98)
    sla_violations = sum(1 for w in predictions if w['response_time'] > sla_deadline) / len(predictions)
    avg_response = np.mean([w['response_time'] for w in predictions])
    avg_energy = np.mean([w['energy'] for w in predictions])
    ips = [w['ips'] for w in predictions]
    fairness = (np.sum(ips)**2) / (len(ips) * np.sum(np.array(ips)**2))
    return {'SLA_Violation_Rate': sla_violations, 'Avg_Response_Time': avg_response, 'Avg_Energy': avg_energy, 'Fairness': fairness}

Common pitfalls

• SLA deadline is dynamically derived per application from the Closure baseline's 98th percentile response time, not a fixed global threshold.
• Workload arrival follows a Poisson process (λ=1.2 physical, λ=5 simulated) normalized by computational requirements, which must be replicated for fair comparison.
• Energy consumption is reported as normalized interval averages, not raw joules, and is combined with response time in the optimization objective but evaluated separately.

Evidence (verbatim from paper)

To compare the QoS, we consider metrics like energy consumption and response time. We also compare the SLA violation rates. The SLA of a workflow is violated if its response time is greater than the deadline. We consider the relative definition of SLA (as in[7]) where the deadline is the 98th percentile response time for the same application (BLAST/Cycles/Montage) on the state-of-the-art baseline Closure.

Citation

@misc{tuli2021mcds,
  title={MCDS: AI Augmented Workflow Scheduling in Mobile Edge Cloud Computing Systems},
  author={Tuli et al. (2021)},
  year={2021},
  note={arXiv:2112.07269}
}

• arXiv: 2112.07269