PaperlayerView exact revision on arXiv

Source-linked AI summary

On the Impact of Degradation-Balanced Scheduling in Manycore Systems

Maria Pantazi-Kypraiou, George Goudroumanis, George Floros, Anuj Pathania, George Stamoulis

arXiv:2610.08190v1cs.AR

TL;DR

Candidate Rvalues can be compressed or tied at scheduling time, leaving the value of current-state reliability metrics uncertain. The paper proposes a scheduler that predicts candidate-assignment effects and balances degradation-related quantities; compared with current Rvalue-temp, it improves minimum Rvalue and several imbalance measures while preserving task completion.

  • Problem

    Candidate Rvalues were often compressed or tied in evaluated decisions, leaving it unclear whether current reliability metrics distinguish candidate cores at decision time.

    Source: A. Limitations of Current-Rvalue… · p. 3

  • Method

    Reliability-Balanced scheduling estimates candidate-assignment consequences and balances EM exposure, stress, active time, and task distribution while preserving a reliability guard.

    Source: §I · p. 1

  • Results

    4.4% higher minimum Rvalue than current Rvalue-temp; RB also reduces Rvalue, exposure, stress, and active-time standard deviations while preserving the same number of completed tasks.

    Source: B. Scheduler Comparison · p. 4

  • Takeaways & Limitations

    RB is about 14× faster than Horizon Exposure in the implementation, though both predictive methods are slower than the myopic baselines.

    Source: C. Discussion · p. 4

    In evaluated decisions, candidate Rvalues were often compressed, with a median range of zero and 96.9% effectively tied in Rvalue.

    Source: A. Limitations of Current-Rvalue… · p. 3

Abstract

from arXiv · show

Electromigration degradation in manycore systems depends strongly on how runtime workload is distributed across cores. Task placement affects power, temperature, and current density, which in turn shape the spatial distribution of degradation over time. However, a scheduler that ranks cores only by their current reliability state may not always distinguish among candidate cores, especially when reliability values are close at decision time. In this case, scheduling decisions can be driven by secondary factors rather than by long-term degradation balance. This paper studies degradation-balanced scheduling and proposes a Reliability-Balanced (RB) policy that predicts the full-task effect of each candidate assignment. RB penalizes added imbalance in electromigration exposure, stress, active time, and task count while guarding the predicted Rvalue floor. On randomized continuous mixed workloads, the proposed scheduler preserves completed tasks while improving average Rvalue by 0.37% and minimum Rvalue by 4.35%. It reduces Rvalue, exposure, stress, and active-time standard deviation by 10.84%, 9.39%, 8.35%, and 16.38%, respectively.

I. INTRODUCTION

Electromigration degradation motivates reliability-aware scheduling, but whether runtime reliability metrics distinguish candidate cores at decision time remains underexplored. The paper introduces a degradation-balancing scheduler and evaluates it in a fixed-core manycore scheduling environment.

I. INTRODUCTION

  • Electromigration is a major degradation mechanism driven by current flow and elevated temperature, motivating runtime scheduling to manage how degradation accumulates across the chip.It can increase interconnect resistance, create voids, and ultimately cause timing violations or connectivity failures.Source: §I · p. 1
  • Whether runtime reliability metrics provide enough decision-time information to meaningfully influence task placement remains largely unexplored.Prior physics-based stress models have been incorporated into system-level reliability management, scheduling, and lifetime optimization frameworks.Source: §I · p. 1
  • Reliability-Balanced scheduling predicts candidate-assignment consequences and balances EM exposure, stress, active execution time, and task distribution while preserving a reliability guard.The evaluation uses continuous mixed workloads and reports improved balance and reliability while preserving task completion.Source: §I · p. 1

A. Scheduling Environment

  • The environment models an abstract homogeneous 16-core system in which each core runs at most one non-preemptive task, and policy differences come from core selection rather than task reordering.Tasks execute until completion, with task duration independent of the selected core.Source: A. Scheduling Environment · pp. 1–2

B. Task and Core Model

The model tracks task and per-core states, updates temperature and electromigration stress, and converts accumulated exposure into a bounded reliability indicator. Its stress update is a reduced-order, per-core approximation rather than a full spatial solution.

B. Task and Core Model

  • Each task has an arrival time, duration, and power demand; each core tracks temperature, reliability, EM exposure, stress, active time, and assigned-task count.Balance-oriented schedulers use these per-core distributions to avoid concentrating work and degradation on a small subset of cores.Source: B. Task and Core Model · p. 2

C. Thermal and Reliability Updates

  • The simulator updates temperature after assignment using a compact RC-style thermal model with effective thermal capacitance and resistance.Its scheduler-level calibration sets a nominal ambient temperature of 40°C and a thermal time constant of about 12 timesteps.Source: C. Thermal and Reliability Updates · p. 2
  • Electromigration degradation is represented by a per-core aggregate stress state, not a full spatial solution of the Korhonen stress-diffusion equation.The model approximates the spatial stress field and retains the assignment-dependent driving term involving current density and temperature-dependent diffusivity.Source: C. Thermal and Reliability Updates · p. 2
  • The stress-rate law uses a temperature-dependent Arrhenius diffusivity normalized around a reference temperature to model task-driven stress growth.The exponential factor is the normalized Arrhenius diffusivity expressed around the reference temperature, not an independent heuristic.Source: C. Thermal and Reliability Updates · p. 2
  • For positive model parameters, stress increase is monotone in current density and temperature, consistent with the scheduler-level assumption that hotter, higher-current assignments accumulate stress faster.The simulator applies a forward-Euler discretization to update stress over each timestep.Source: C. Thermal and Reliability Updates · p. 2
  • The simulator maps stress-based time-to-failure to a Weibull lifetime, accumulates normalized exposure, and calculates an Rvalue that starts at 1 and decreases as exposure accumulates.The Rvalue is used for core validity checks and final reliability evaluation.Source: C. Thermal and Reliability Updates · pp. 2–3

1) Rvalue-Temperature Policy:

The Rvalue-temperature policy scores valid cores using current reliability and temperature. Its current-state scoring can be dominated by temperature or tie-breaking when candidate Rvalues are close.

1) Rvalue-Temperature Policy:

  • The policy selects the valid core with the largest current-state score, subtracting a heuristic temperature penalty with λ = 0.01.It scans candidate cores and has O(C) selection cost once the C valid candidates are available.Source: 1) Rvalue-Temperature Policy · p. 3
  • When candidate Rvalues are compressed or tied, temperature and tie-breaking can dominate the decision despite the policy’s reliability-aware framing.The policy relies only on the current reliability state.Source: 1) Rvalue-Temperature Policy · p. 3
  • The policy evaluates a scalar score per candidate, giving O(N) worst-case selection cost or O(C) when valid candidates are already available.Source: 1) Rvalue-Temperature Policy · p. 3

2) Horizon Exposure Policy:

Horizon Exposure predicts exposure growth by rolling out candidate assignments before choosing a core. It uses the smallest predicted cumulative exposure growth as its selection criterion, but does not directly balance exposure, stress, or workload across cores.

2) Horizon Exposure Policy:

  • For each valid core, Horizon Exposure copies the simulator state, assigns the waiting task, then rolls out up to K = 8 subsequent assignments using a greedy minimum-exposure-growth rule.The rollout spans H simulated timesteps, depending on task durations and arrival times.Source: 2) Horizon Exposure Policy · p. 3
  • The selected core minimizes predicted cumulative exposure growth, comparing each core’s current exposure with its predicted exposure after the rollout.The simulator cost is O(CHN), because each candidate rollout updates all cores at each simulated timestep.Source: 2) Horizon Exposure Policy · p. 3
  • Exposure growth is a global consumption measure and does not explicitly penalize uneven exposure, stress, or workload distribution between cores.Source: 2) Horizon Exposure Policy · p. 3

3) Proposed Reliability-Balanced (RB) Policy:

The Reliability-Balanced policy predicts each candidate’s post-task state and scores assignments for their effects on degradation and workload balance. Its reported configuration also guards the predicted minimum Rvalue.

3) Proposed Reliability-Balanced (RB) Policy:

IV. EXPERIMENTAL EVALUATION

The evaluation compares four scheduling policies on randomized continuous-arrival workloads using an abstract 16-core platform. Since the policies complete the same number of tasks on average, the comparison focuses on work and degradation distribution rather than throughput.

IV. EXPERIMENTAL EVALUATION

  • The main evaluation uses an abstract 16-core manycore platform and 20 randomized continuous-arrival workloads with overlapping tasks and mixed duration and power demand.Source: §IV · p. 3
  • The evaluation compares Round Robin, the Rvalue-temperature rule, Horizon Exposure, and the proposed RB policy.Source: §IV · p. 3
  • All policies complete the same number of tasks on average, so the comparison mainly reflects work and degradation distribution rather than throughput differences.Source: §IV · p. 3

A. Limitations of Current-Rvalue Scheduling

Candidate Rvalues are often too similar at decision time to distinguish cores, and temperature often separates candidates under the same score. This motivates balancing degradation-related quantities across cores rather than relying only on current Rvalue.

A. Limitations of Current-Rvalue Scheduling

B. Scheduler Comparison

RB improves reliability and degradation balance over current Rvalue-temp while preserving task completion; Horizon Exposure also improves several metrics but does not explicitly optimize balance.

B. Scheduler Comparison

  • Current Rvalue-temp’s baseline mean and minimum Rvalues are 0.922 and 0.684, respectively.Its Rvalue, exposure, and stress standard deviations are 0.090, 0.156, and 40.401, respectively.Source: B. Scheduler Comparison · p. 3
  • Horizon Exposure improves several Rvalue and exposure metrics, but exposure growth alone does not explicitly optimize degradation balance and its rollout is substantially more expensive.The cost qualifier applies to the current implementation.Source: B. Scheduler Comparison · p. 4
  • Compared with current Rvalue-temp, RB improves minimum Rvalue by 4.4%, reduces Rvalue standard deviation by 10.8%, and reduces maximum degradation difference by 10.3%.The comparison is on the continuous mixed workload; Table I reports relative improvement over current Rvalue-temp.Source: B. Scheduler Comparison · p. 4, Table I · p. 4
  • RB reduces exposure standard deviation by 9.4%, stress standard deviation by 8.3%, and active-time standard deviation by 16.4% versus current Rvalue-temp.It preserves the same number of completed tasks.Source: B. Scheduler Comparison · p. 4, Table I · p. 4
  • Figure 1’s per-core exposure and active-timestep distributions show reduced degradation concentration and workload concentration with RB.The upper panel shows final EM exposure, and the lower panel shows active timesteps; lower dispersion indicates less concentration.Source: B. Scheduler Comparison · p. 4, Figure 1 · p. 4

C. Discussion

The conclusion reports that RB improves degradation distribution and reliability without affecting task completion. The discussion distinguishes these balance gains from runtime cost and notes that real-system scheduling overhead remains future work.

C. Discussion

  • RB is about 14× faster than Horizon Exposure in the reported implementation, though both predictive methods are slower than the myopic baselines.Round Robin and Rvalue-temp have similar measured runtime; Horizon Exposure is most expensive because of rollout simulation.Source: C. Discussion · p. 4
  • The results are framed as degradation-balance improvements, not evidence that any single reliability metric is sufficient for scheduling..Source: C. Discussion · p. 4
  • RB is evaluated as a simulator-level placement heuristic at task dispatch, while real-system scheduling latency and integration overhead remain future work.All policies complete the same number of tasks in the workload.Source: C. Discussion · p. 4

V. CONCLUSION

  • On continuous randomized mixed workloads, RB raises minimum Rvalue by 4.4% and reduces Rvalue, exposure, stress, and active-time imbalance by up to 10.8%, 9.4%, 8.3%, and 16.4%, without affecting task completion.These results compare approaches from current-state reliability metrics to predictive estimation and RB’s balancing of degradation-related quantities with a reliability guard.Source: §V · p. 4
Loading 2610.08190v1…