COMPASS-ABS: Reducing Fragmentation in Shared GPU Clusters for Deep Learning Training Workloads
cs.DC · cs.LG
Quick verdict
COMPASS-ABS proposes a history-free fragmentation metric (SIF) plus an anchor/slot/pool scheduler with a proven SIF≤2/N guarantee under a workload-composition condition; conceptually sound and well-evaluated, but evaluation breadth, migration overhead accounting, and baseline fairness need scrutiny.
Summary
The paper targets fragmentation in shared GPU clusters running deep learning training (DLT) jobs, where intra-node placement constraints strand idle GPUs across partially-filled nodes. The authors first define Scheduler-Induced Fragmentation (SIF): total partial-node share minus an 'inherent' floor computed as the optimal bin-packing partial-node count for the current running-job set, making the metric independent of historical workload distributions. They then design COMPASS-ABS, a scheduler that keeps the cluster inside an 'Anchor-Based Space' (ABS): multi-GPU instances become anchors of width-2/4/8 slots, 1-GPU instances serve as fillers or live in a dedicated pool, and a COMPASS algorithm (Place/Remove/Compact operators with best-fit-drain compaction) runs after every event. Under a Workload Composition Condition (WCC) ensuring enough 1-GPU fillers, they prove SIF≤2/N at all times. Evaluation spans a 4-node physical A100 cluster and simulations on Venus and GFS traces against six baselines, showing lowest AFR, highest AGU, and shortest AJCT, robust to η∈[0.85,0.95], trace-day composition, migration cost 1–10s, and mostly robust when WCC is violated. Limitations: the theory relies on 8-GPU nodes, power-of-two demand dominance, and gang-scheduled homogeneous instances; migration overhead is deferred; generality to heterogeneous/elastic or inference workloads is future work.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- SIF, defined as observed partial-node share minus the optimal-packing inherent floor, measures scheduler-attributable fragmentation without workload history (Sec. 3).
- Under the Workload Composition Condition, COMPASS-ABS keeps SIF≤2/N at all times (Theorem 1, proven in Appendix B).
- COMPASS-ABS attains the lowest fragmentation, highest GPU utilization, and shortest job completion time among seven schedulers on Venus, GFS, and a physical cluster (Sec. 6.1).
- The advantage persists across cluster intensity (η∈0.85–0.95), workload composition shifts, migration cost 1–10s, and most WCC-violating days (Secs. 6.2–6.5).
- The framework extends to multi-panel NVLink domains (NVL72/NVL576) via a hierarchical ABS with per-level WCC (Appendix D, Theorem 2).
Strengths
History-free fragmentation metric with a clean decomposition
SIF (Sec. 3.2) subtracts an inherent floor Φ*(t)—the optimal partial-node count over all feasible packings of currently running jobs—from observed fragmentation, isolating scheduler-attributable avoidable partial nodes without workload history, unlike FGD/CAFGD statistical measures. The avoidable/forced partial-node examples in Sec. 3.1 make the decomposition intuitive.
End-to-end theoretical guarantee tied to a checkable condition
Theorem 1 proves SIF≤2/N whenever the Workload Composition Condition (Eq. 9) holds, with proofs (Appendix B) covering state-space invariance (Prop. 1), slot saturation (Lemma A.2), and pool/slot compaction (Lemmas A.3/A.4). Sec. 6.4 empirically measures WCC coverage ρ on every replayed Venus/GFS day (86.3% of Venus days, 100% of GFS days fully covered) and reports top-1 rates in violating buckets, connecting theory to practice.
Design grounded in production workload structure
Multi-regime evaluation including WCC-violation robustness
Weaknesses
Small physical testbed; simulation dominance
Migration cost treated as a constant parameter only
WCC is strong and its violation analysis is aggregate
Comparison metric set mixes diagnostic and outcome metrics
Text-extraction artifacts cloud the appendices
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Scheduler-Induced Fragmentation (SIF)?
SIF is a fragmentation metric for GPU clusters that subtracts an 'inherent' floor—the minimum number of partially-filled nodes achievable by any packing of currently running jobs—from the observed partial-node share. The remainder isolates fragmentation caused by the scheduler's placement decisions, and unlike FGD/CAFGD it needs no historical workload distribution.
What is the Workload Composition Condition (WCC) in COMPASS-ABS?
WCC requires the number of active 1-GPU instances to be at least n3+3n5+2n6+n7, the total slot vacancy left by fragmentation-unfriendly (3/5/6/7-GPU) anchors. When it holds at every instant, Theorem 1 guarantees the scheduler keeps at most two partially-filled nodes, i.e., SIF≤2/N.
How does COMPASS-ABS compare to FGD and CAFGD?
FGD and CAFGD optimize statistical fragmentation measures that require historical workload distributions and act only at first placement; COMPASS-ABS uses a history-free metric, actively migrates 1-GPU fillers and slots after every event via the COMPASS algorithm, and provides a worst-case SIF bound under WCC, dominating both in the reported simulations.
Does COMPASS-ABS require power-of-two GPU requests?
The strong 2/N guarantee requires the workload to satisfy the Workload Composition Condition, which in practice holds because power-of-two sizes {1,2,4,8} dominate production traces (Table 1: fragmentation-unfriendly sizes are 0.05–0.44% of instances). Non-power-of-two requests are handled by rounding into slots and filling with 1-GPU instances, so the scheduler still operates when the condition is temporarily violated, just without the bound.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–7 are not analyzable from extracted text; quantitative claims rely on the prose numbers in Sec. 6.
- The embedded text of Appendices A and C is heavily corrupted by PDF extraction (garbled sentences), so Proposition 3's tightness construction and Algorithm 2 could not be fully verified.
- Only abstract plus extracted body text was available; no review of supplementary materials or the anonymous code repository was performed.
- Simulation confidence intervals and per-day variance are not extractable, so effect-size robustness cannot be independently confirmed.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.