Source-linked AI summary
PaFeDSE: Pareto-Feedback-Guided Microarchitecture Design Space Exploration for Efficient Pareto Search
Joonyong Park, Hyojeong Lee, Sejeong Yang, Joo-Young Kim
TL;DR
Existing ML-based microarchitecture DSE wastes simulations on Pareto-miss candidates and repeatedly processes low-priority designs. PaFeDSE combines Pareto-feedback-guided selection with adaptive candidate filtering; on SPEC CPU 2017, it reaches target hypervolume levels with up to 49.0% fewer simulations than the strongest prior baseline and reduces exploration overhead 7.1×.
Problem
Existing ML-based DSE methods still spend costly simulations on Pareto misses and repeatedly process candidates that remain low priority across iterations.
Source: §1 · p. 2Method
PaFeDSE combines HitFormer’s PPA and hit-score prediction, MissGraph’s local penalties around Pareto misses, and TCFilter’s inter-iteration candidate filtering.
Source: §1 · p. 2Results
Up to 49.0% fewer simulations reach target hypervolume levels than the prior state-of-the-art method, while adaptive filtering reduces exploration overhead 7.1× across SPEC CPU 2017 workloads.
Source: §1 · p. 2Takeaways & Limitations
PaFeDSE achieves the highest average final normalized hypervolume under the same simulation budget, identifying high-quality Pareto solutions across SPEC CPU 2017 workloads.
Source: §6 · p. 7Because the true Pareto frontier is computationally infeasible to obtain, target hypervolumes are defined as 90%, 95%, and 99% of the best final HV among compared methods under a 200-simulation budget.
Source: §4.2 · p. 6
Abstract
from arXiv · showhide
Efficient microarchitecture design space exploration (DSE) is critical for optimizing performance, power, and area trade-offs in modern out-of-order processors. While machine-learning-based DSE significantly reduces the cost of exhaustively evaluating large design spaces, existing methods still waste expensive simulations on Pareto-miss candidates that do not improve the Pareto frontier and repeatedly process redundant candidates during iterative exploration. In this paper, we propose PaFeDSE, a Pareto-feedback-guided microarchitecture DSE framework for efficient Pareto search. PaFeDSE leverages Pareto-hit feedback to prioritize candidates likely to improve the frontier and Pareto-miss feedback to suppress exploration of unpromising local regions, while filtering redundant candidates that remain low priority across iterations. Evaluated on SPEC CPU 2017 workloads, PaFeDSE requires up to 49.0% fewer simulations than the strongest prior baseline to reach target hypervolume levels, while achieving a 7.1x reduction in exploration overhead through adaptive candidate filtering.
1 Introduction
Existing ML-based microarchitecture DSE still spends costly simulations on Pareto misses and repeatedly processes low-priority candidates. PaFeDSE uses Pareto-hit and Pareto-miss feedback plus inter-iteration consistency to improve search efficiency.
1 Introduction
- MoDSE’s Pareto-miss rate reaches 85.47% on evaluated SPEC CPU 2017 workloads, while existing methods do not use hit-or-miss outcomes to guide later candidate selection.A Pareto hit is a selected candidate whose evaluated PPA increases observed hypervolume by more than ε; otherwise it is a miss.Source: §1 · p. 2
- Low-rank candidates often remain low priority across iterations, yet existing methods repeatedly process all unevaluated candidates, adding overhead unlikely to change acquisition outcomes.The paper identifies high rank-bin retention in low-rank bins as evidence of this persistence.Source: §1 · p. 2
- HitFormer predicts PPA and Pareto-hit scores, MissGraph penalizes candidates near prior misses, and TCFilter removes persistently low-priority candidates.Together, these components use Pareto feedback to guide selection and inter-iteration consistency to reduce redundant processing.Source: §1 · p. 2
- Up to 49.0% fewer simulations reach target hypervolume levels than the prior state-of-the-art method, while adaptive filtering reduces exploration overhead 7.1×.Evaluations use SPEC CPU 2017 workloads.Source: §1 · p. 2
2 Background
Microarchitecture DSE searches costly-to-evaluate design spaces for Pareto-optimal performance, power, and area trade-offs. The background defines the Pareto set and hypervolume measures used to describe search quality and candidate improvement.
2.1 DSE Problem Formulation
- DSE maps each microarchitecture configuration to an objective vector, then uses surrogate-assisted iterative exploration to evaluate a small fraction of the design space.At each iteration, a surrogate estimates unevaluated candidates and an acquisition function selects one likely to improve the observed Pareto frontier.Source: §2.1 · p. 2
2.2 Pareto Optimality and Hypervolume
- A design point is Pareto-optimal when no other point dominates it, meaning the Pareto set contains designs not dominated by any point in the design space.Dominance requires being no worse in every objective and strictly better in at least one.Source: §2.2 · p. 2
- Hypervolume measures the objective-space volume dominated by the Pareto set relative to a reference point, and HVI measures the increase from adding a candidate.In minimization, the reference point is worse than Pareto points in every objective and bounds the dominated region.Source: §2.2 · p. 2
- PaFeDSE’s iterative procedure scores active candidates using predicted HVI, estimated hit probability, and a miss-graph factor, then simulates the top candidate and filters the pool.The procedure updates the observed Pareto set and miss graph between iterations.Source: §2.2 · p. 3
3 PaFeDSE Framework
PaFeDSE combines Pareto-hit-aware scoring, locality-aware Pareto-miss penalties, and temporal filtering in an iterative search that updates the observed Pareto set. Its components aim to prioritize promising candidates while reducing repeated processing of persistently low-priority candidates.
3.1 Overview of PaFeDSE
- PaFeDSE initializes an active candidate pool, then iteratively selects and simulates candidates, updating the observed Pareto set, MissGraph, and pool.The initial evaluated set is chosen with sensitivity-aware sampling.Source: §3.1 · p. 3
3.2 Pareto-Hit-Aware Transformer Surrogate
- HitFormer jointly predicts PPA and Pareto-hit likelihood, conditioning candidate representations on tokens from the current observed Pareto set.Its acquisition score combines predicted HVI with hit likelihood; training uses leave-one-out pseudo-DSE samples and upweights high-HVI Pareto-miss hard negatives.Source: §3.2 · pp. 3–4
3.3 Locality-Aware Pareto-Miss Penalty Graph
- MissGraph penalizes candidates near observed Pareto misses using a sensitivity-weighted design distance and a coarse-to-fine neighborhood schedule.The graph factor is non-positive in penalty space and scales the acquisition score down for penalized candidates.Source: §3.3 · p. 4
3.4 Temporal-Consistency-Guided Filtering
- TCFilter removes only candidates repeatedly ranked in the bottom k% across a window, then periodically refreshes the active pool to reconsider filtered candidates.It also enforces a minimum active-set size.Source: §3.4 · pp. 4–5
4 Evaluation
Across SPEC CPU 2017 workloads, PaFeDSE reaches target hypervolume levels with fewer simulations than prior methods and achieves the highest average final normalized HV under the same budget. Its ablations and detailed analyses show how Pareto feedback and candidate filtering affect simulation efficiency, prediction, and exploration overhead.
4.1 Evaluation Setup
- The evaluation uses gem5 and McPAT on a 3,359,232-point design space, with SPEC CPU 2017 workloads and fixed L2-cache and DRAM configurations.Baselines use an initial set of 28; experiments run on a 36-core Xeon Platinum 8452Y CPU and an RTX 6000 Ada Generation GPU.Source: §4.1 · p. 5
4.2 Overall Pareto Quality and Search Efficiency
- Up to 49.0% fewer simulations than MoDSE were needed to reach the 99% target HV; PaFeDSE also achieved the highest average final normalized HV across workloads under the 200-simulation budget.On average, it reached the 90%, 95%, and 99% target HVs in 32.9, 43.3, and 118.8 simulations, respectively.Source: §4.2 · p. 6
4.3 Ablation Study
- Adding the Pareto-hit score and MissGraph reduced simulations to the 98% target HV from 90.0 to 81.7 and then 69.3; TCFilter cut average exploration time per iteration from 374.19 s to 52.94 s without degrading final HV.The ablation incrementally adds the Pareto-hit score, MissGraph, and TCFilter to an HVI-only HitFormer.Source: §4.3 · p. 6
4.4 Detailed Analysis of PaFeDSE
- PaFeDSE's Pareto-hit score outperformed random-ranking and constant-probability baselines across AP, ROC-AUC, and Brier score on candidates selected during actual DSE runs.The study also evaluates PPA prediction MAPE over points selected during each method's DSE process.Source: §4.4 · p. 6
- The average active candidate ratio fell to 1.82%, while PaFeDSE's average PPA MAPE was 0.731%, 30.2% lower than MoDSE's.The active ratio is |A_t|/|U_t|; TCFilter filters out the vast majority of candidates.Source: §4.4 · p. 7
4.5 Time Analysis of PaFeDSE
- With parallel SimPoint execution, evaluation time fell from 3962.7 s to 207.9 s per iteration, and overall execution time including exploration fell from 4357.5 s to 602.7 s without TCFilter.The evaluation totals include gem5 simulation and McPAT estimation.Source: §4.5 · p. 7
- Under parallel evaluation, TCFilter reduced exploration time from 394.8 s to 60.4 s and total per-iteration execution time from 602.7 s to 268.3 s.Exploration time includes model training, inference, and acquisition scoring.Source: §4.3 · p. 6, §4.5 · p. 7
- TCFilter reduced inference and acquisition-scoring time by shrinking the candidate pool, while combined model-training time remained nearly unchanged.Three-dimensional HVI computation for active candidates is a substantial part of acquisition scoring.Source: §4.5 · p. 7
5 Related Work
Prior ML-based microarchitecture DSE methods use diverse learning and acquisition strategies, while PaFeDSE explicitly uses Pareto outcomes and temporal consistency to guide exploration.
Related ML-based DSE
- ActBoost uses active learning, BOOM-Explorer combines GP surrogate modeling with EHVI acquisition, and MoDSE uses Pareto-aware surrogate training, HVI selection, and uniformity-aware candidate selection.These methods are described as approaches to reducing expensive design evaluations.Source: §5 · p. 7
Related learning-based approaches
PaFeDSE
- PaFeDSE differs by using Pareto-hit and Pareto-miss outcomes to guide candidate selection and temporal consistency to reduce redundant inference and acquisition scoring.Source: §5 · p. 7
6 Conclusion
PaFeDSE combines HitFormer, MissGraph, and TCFilter for CPU microarchitecture DSE. Across SPEC CPU 2017 workloads, it reports fewer simulations, lower PPA prediction error, and less exploration overhead, alongside the highest average final normalized HV under the same simulation budget.
Conclusion
- Across SPEC CPU 2017 workloads, PaFeDSE required up to 49.0% fewer simulations to reach high-HV regions and reduced PPA prediction error by 30.2% over the strongest prior baseline.Source: §6 · p. 7
- Adaptive filtering achieved a 7.1× reduction in exploration overhead, and PaFeDSE achieved the highest average final normalized HV under the same simulation budget.Source: §6 · p. 7
- The paper concludes that Pareto feedback and temporal candidate consistency are an effective direction for improving simulation and candidate-processing efficiency in microarchitecture DSE.Source: §6 · p. 7