Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Decouple and Decompose: Scaling Resource Allocation with DeDe

T0 review · 4 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read DeDe splits large resource-allocation problems into many tiny parallel subproblems by decoupling resource and demand constraints, with reported near-optimal quality and large speedups.

desk verdict The convex ADMM decomposition is a genuine contribution; the integer/nonconvex claims and the load-balancing formulation overreach. read the letter →

arxiv 2412.11447 v3 pith:6FIATDP7 submitted 2024-12-16 cs.DC

Zhiying Xu , Minlan Yu , Francis Y. Yan This is my paper · ORCID
classification cs.DC
keywords resourceallocationADMMconstraintdecompositionparalleloptimizationclusterschedulingtrafficengineeringloadbalancingseparable
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

DeDe claims that most real-world resource allocation problems share a separable structure: the objective is a sum over per-resource and per-demand utilities, and every constraint applies to a single resource or a single demand. The central claim is that this structure can be exploited by a decouple-and-decompose reformulation: introduce an auxiliary copy $z$ of the allocation matrix, move demand constraints onto $z$, add $x=z$, and run ADMM. This turns one large entangled optimization into many small per-resource and per-demand subproblems that can be solved in parallel with off-the-shelf solvers, producing near-optimal allocations. If correct, DeDe is a domain-agnostic accelerator for cluster scheduling, traffic engineering, and load balancing, with reported speedups of 2.2--7.6$\times$ and allocation-quality gains of 5.3--12.6% over the previous granular method.

What carries the argument

The central mechanism is a two-block ADMM splitting of the allocation matrix. DeDe replaces the allocation matrix $x$ with a duplicate $z$, moves all demand constraints and demand-dependent objective terms onto $z$, and couples the blocks with $x - z = 0$. The augmented Lagrangian with penalty $\rho$ and scaled dual variables $\alpha$, $\beta$, $\lambda$ is then minimized alternately: for fixed $z$, the $x$-update separates into $n$ per-resource subproblems (Equation 8); for fixed $x$, the $z$-update separates into $m$ per-demand subproblems (Equation 9). Because every subproblem retains access to the full set of resources or demands, the decomposition does not shrink the feasible region the way subset partitioning does.

What would settle it

Run DeDe on a small mixed-integer load-balancing instance (for example, 8 servers and 16 shards) whose exact optimum can be verified by exhaustive search; if DeDe, across a sweep of the penalty parameter, returns objective values systematically worse than that known optimum on instances where the continuous relaxation is tight, the near-optimal-integer claim fails.

Watch

Extended reading notes

Core claim

The discovery is that the coupling between resource constraints and demand constraints, rather than the raw number of variables, is the main obstacle to parallelizing resource allocation. For any problem of the form $\min \sum_i f_i(x_{i*}) + \sum_j g_j(x_{*j})$ subject to $R_i x_{i*} = r_i$ and $D_j x_{*j} = d_j$, DeDe duplicates the allocation matrix, rewriting the demand-side constraints and utilities on the copy $z$ while adding $x - z = 0$. The augmented Lagrangian for this equivalent problem is then minimized by ADMM: an $x$-update that separates into $n$ independent per-resource subproblems and a $z$-update that separates into $m$ independent per-demand subproblems, with multiplier updates in between. The paper argues this preserves the optimum for convex problems and, empirically, yields near-optimal allocations for the non-convex, integer load-balancing case while cutting solving time by large factors relative to exact solvers and by 2.2--7.6$\times$ relative to the prior subset-splitting method.

Load-bearing premise

For problems with whole-number or yes/no choices, DeDe assumes that repeatedly projecting ADMM's real-valued steps onto the allowed discrete values still lands near the best solution; the paper offers empirical evidence and special-case theory, but no proof for the general integer case.

Editorial extensions

If this is right

  • Cluster scheduling, traffic engineering, and load balancing problems that fit the separable form can be solved in seconds rather than minutes to hours, with allocation quality at or above what exact solvers achieve within their time limits.
  • The number of parallel subproblems grows with the number of resources and demands, so adding CPU cores transfers to near-linear speedup until inter-process overhead and stragglers dominate.
  • Because each subproblem still sees all resources or all demands, DeDe's solution quality degrades far less than subset-splitting methods when demands are non-granular or resources are not interchangeable.
  • Warm-starting DeDe with a previously computed or learned allocation roughly halves the time to a good solution, making the method natural for repeated optimization intervals.
  • The decomposition is domain-agnostic: any new allocation problem expressible in the separable form can be written into the package without per-domain solver logic.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the paper's survey of separable structure generalizes, the same template should apply to other allocation settings the paper catalogues but does not evaluate, such as intercloud brokering, optical-path wavelength assignment, and electricity pricing; these are natural testbeds.
  • The real 64-core speedup saturates near 18$\times$ because of cache contention and stragglers, so dynamic work-stealing or running subproblems across machines is a plausible engineering extension the paper does not pursue.
  • The integer-variable success is empirical; stress-testing DeDe against known-optimal mixed-integer benchmarks would map where the near-optimality claim holds and where it breaks.
  • Combining DeDe's decomposition with a learned coarse initializer (as the paper tests with one learning-based baseline) suggests a general recipe: leverage machine learning to propose a start and ADMM splitting to refine it.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes DeDe, a framework for large-scale resource allocation problems that have separable objective and constraint structure. DeDe introduces an auxiliary variable z duplicating the allocation matrix x, moves demand-side constraints and objective terms onto z, and applies scaled-form ADMM to alternate between independent per-resource updates (Eq. 8) and per-demand updates (Eq. 9). The paper claims this decouple-and-decompose approach yields near-optimal solutions without compromising quality, and presents a Python package plus experiments on cluster scheduling, traffic engineering, and load balancing, reporting speedups and quality improvements over POP and other baselines.

Significance. For continuous convex separable problems, the ADMM-based decomposition is standard and the derivation in Section 3 is sound; the evaluation shows meaningful empirical speedups, and the public release of a working parallel implementation is a concrete systems contribution. The central weakness is that the paper extends the claim to integer and non-convex problems (load balancing) using an unproven projection heuristic, and the abstract's 'without compromising solution quality' assertion is not supported for that class. If the claims are narrowed to continuous convex problems and the integer extension is presented as a heuristic with appropriate caveats, the paper would be a useful contribution to scalable resource allocation.

major comments (4)
  1. [§4.1, Eqs. (8)–(9)] The claim that DeDe 'should effectively handle boolean and integer variables by projecting real-valued solutions onto the appropriate domains' is not supported by the cited references. References [39,57,61] do not establish convergence or solution-quality guarantees for the two-block ADMM scheme in Eqs. (8)–(9) when subproblems contain mixed-integer variables, nor for the inexact subproblem solves used in practice. Section 4.2 concedes that DeDe 'may fail to reach the optimal solution' in non-convex settings, which directly conflicts with the introduction's claim of 'without compromising solution quality.' Since the load-balancing evaluation in §5.3 relies on binary variables, this gap is load-bearing and should be addressed by either restricting the generality claims or providing concrete convergence/quality analysis or substantially more evidence for the integer class.
  2. [§5.3, Eq. (8)] The load-balancing formulation is incomplete and, as stated, degenerate. The continuous variables x_ij and binary variables x'_ij are not linked by any constraint such as x_ij <= x'_ij or x_ij > 0 => x'_ij = 1. Without such a coupling, the objective and memory constraints depend only on x', so setting x' = 0 is always optimal for any feasible x, and the model does not actually minimize shard movements. With the intended coupling, each per-server subproblem becomes an MILP with m binary variables, whose exact solution is NP-hard, and the paper provides no convergence or approximation guarantee for ADMM on this class. This undermines the load-balancing results in Figure 8 and the related claims of higher allocation quality.
  3. [§3.2, complexity comparison] The complexity statement that decomposition reduces the original O((n m)^2.373) cost to O(n m^2.373) is explicitly conditional on ADMM 'converging within a constant number of iterations,' but no such constant-iteration bound is provided or cited for the general separable problem class. The iteration count depends on the penalty parameter rho, stopping tolerances, and problem conditioning, and these are free parameters in the implementation. The theoretical speedup claim should therefore be presented as an empirical observation, not a worst-case complexity result, or it should be backed by a convergence-rate analysis.
  4. [§7.1.3 and Fig. 8] The load-balancing comparison does not substantiate the 'higher-quality allocations' claim. The 'Exact sol.' baseline solves each scheduling round independently, and the text admits it 'does not produce the minimum possible shard movements because optimizing each scheduling round independently does not guarantee an optimal solution across the entire series.' No optimality gap or constraint-violation statistics are reported for any method. Since DeDe's average of 20.1 shard movements is better than Exact sol.'s 20.9, the evaluation needs a proper sequence-level baseline or a per-round gap analysis before the quality claim can be accepted.
minor comments (4)
  1. [§7.1.1, proportional fairness] In the proportional-fairness experiment, Exact sol. is reported to fail to reach optimality even after five hours, and several methods therefore achieve normalized fairness scores above 1. The paper should clarify that the normalization baseline is not an exact optimum and should report absolute objective values or optimality gaps.
  2. [§4.2, limitations] The discussion of 'limited parallelism' would benefit from a concrete example or bound quantifying how aggregation reduces parallelism, rather than only a qualitative statement.
  3. [§5.2, traffic engineering] The sentence 'the objectives are expressed as either a sum over per-demand utilities or as a maximum over per-link utilities' is slightly imprecise because the maximum over link utilization is not a sum of per-link utilities; the connection to the augmented Lagrangian is clear but the wording could be tightened.
  4. [§7.1.1, cluster scheduling] The comparison of DeDe and DeDe* would be easier to interpret if the paper consistently distinguished end-to-end wall-clock time from solver-only time in every figure, since DeDe* explicitly excludes compilation, solving-unpacking, and scheduling overheads.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: DeDe's ADMM decomposition is an algebraic reformulation, and its speedups and quality are measured against external baselines; the nonconvex/integer gap is an evidence limitation, not a circular reduction.

full rationale

DeDe's derivation is self-contained in the relevant sense. The paper starts from a separable form (Eqs. 1-3), introduces z=x in Eq. (4) as an exact equivalence, and then applies standard two-block ADMM; the per-resource and per-demand subproblems in Eqs. (8)-(9) are obtained by separating the augmented Lagrangian across disjoint indices, so the decomposition is algebraic rather than assumed. Convex convergence is imported from classical external results (Boyd et al.; Gabay-Mercier; Glowinski-Marroco), not from the authors' own prior work, and the reported speedups and quality numbers are empirically measured against commercial solvers and independent baselines (Exact sol., POP, Gandiva, E-Store, Pinning, Teal). The only self-citation, Teal, appears as a comparison baseline and an optional warm-start initializer; it is not load-bearing for the central claim that the separable structure permits decoupling. The closest concern—the Sec. 4.1 assertion that ADMM with integer projection will handle boolean and integer variables—is an unproven evidence gap for a nonconvex class, and the paper itself concedes in Sec. 4.2 that 'DEDE may fail to reach the optimal solution.' That is a correctness/robustness limitation, not a circular reduction of a prediction to its input. The potential degeneracy in the load-balancing MILP due to the missing explicit x/x' coupling is likewise a modeling issue, not a circular derivation. Thus no circular step can be exhibited, and the score is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central method is standard ADMM variable-splitting; the paper's contribution is the observation of separable structure plus its system implementation. The ledger is dominated by domain assumptions (separability, linearity, integer projection) and three unreported or tuned numerical parameters; no genuinely new entities are introduced.

free parameters (3)
  • ADMM penalty parameter rho = not reported
    Controls convergence speed and final solution quality; the paper states iteration count depends on it but does not report values or tuning procedure.
  • ADMM stopping tolerance = not reported
    No stopping criterion is given for terminating the ADMM loop; this affects both quality and runtime.
  • Load-balancing tolerance epsilon = 0.1
    Set to 0.1 to ensure feasibility in the load balancing evaluation (Sec. 7.1.3); not part of the general framework.
assumptions (5)
  • standard math Two-block ADMM converges to an optimal solution for convex problems
    Relied on in Sec. 3.1 to claim optimality for convex separable problems; standard from Boyd et al. [8].
  • domain assumption Real-world resource allocation problems can be written in the separable form of Eqs. (1)-(3)
    Central premise of Sec. 2 and Sec. 4.1; the paper supports it with Table 1 but does not give a formal characterization or negative results.
  • domain assumption All constraints of interest are linear
    Stated in Sec. 4.1; nonlinear constraints destroy convexity of the augmented Lagrangian per Sec. 4.2.
  • ad hoc to paper Projecting ADMM iterates onto integer or binary domains yields high-quality solutions
    Invoked in Sec. 4.1 for boolean and integer variables (load balancing) with only empirical citations; no proof for this problem class.
  • ad hoc to paper ADMM converges within a constant number of iterations for the complexity comparison
    Used in Sec. 3.2 to claim O(n m^2.373) vs O((nm)^2.373); the paper itself notes the iteration count depends on rho.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Decouple and Decompose: Scaling Resource Allocation with DeDe." pith.science (2026). https://pith.science/paper/6FIATDP7

@misc{pith2026241211447,
  author       = {Pith},
  title        = {Pith review of: Decouple and Decompose: Scaling Resource Allocation with DeDe},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6FIATDP7}},
  note         = {Machine review of arXiv:2412.11447}
}
read the original abstract

Efficient resource allocation is essential in cloud systems to facilitate resource sharing among tenants. However, the growing scale of these optimization problems have outpaced commercial solvers commonly employed in production. To accelerate resource allocation, prior approaches either customize solutions for narrow domains or impose workload-specific assumptions. In this work, we revisit real-world resource allocation problems and uncover a common underlying structure: the vast majority of these problems are inherently separable, i.e., they optimize the aggregate utility of individual resource and demand allocations, under separate constraints for each resource and each demand. Building on this observation, we develop DeDe, a scalable and theoretically rooted optimization framework for large-scale resource allocation. At the core of DeDe is a decouple-and-decompose approach: it decouples entangled resource and demand constraints and thereby decomposes the overall optimization into alternating per-resource and per-demand subproblems that can be solved efficiently and in parallel. We have implemented and released DeDe as a Python package with a familiar modeling interface. Our experiments on three representative resource allocation tasks -- cluster scheduling, traffic engineering, and load balancing -- demonstrate that DeDe delivers significant speedups while generating higher-quality allocations.

Figures

Figures reproduced from arXiv: 2412.11447 by the authors.

Figure 1
Figure 1. Overview of DEDE. These problems may involve millions of variables and take tens of minutes or even hours to solve, whereas fast allocation is necessary to meet service-level objectives. This “scalability crisis” has spurred a flurry of studies [1, 43, 48, 65] that trade off solution quality for computational speed, using heuristics, approximate algorithms, or machine learning. Nevertheless, these approaches are typ… view at source ↗
Figure 2
Figure 2. DEDE decouples constraints using an auxiliary variable, thereby decomposing optimization into subproblems. possible—the optimal allocation of a set of large language model inference jobs (e.g., GPT-4, Llama 3, and DeepSeek￾V3) on a heterogeneous GPU cluster composed of various hardware types (e.g., Nvidia H100, A100, and V100). Jobs may divide their execution time across multiple GPU types, and due to differences in… view at source ↗
Figure 3
Figure 3. Toy scenario of the job scheduling example. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (6 more)
Figure 5
Figure 5. Figure 5: Results of the cluster scheduling variant that maxi [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Results of the traffic engineering problem variant [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 8
Figure 8. Figure 8: Results of the load balancing problem where the [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Normalized satisfied demand in traffic engineering against granularity, temporal, and spatial changes. [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Micro-benchmarks with the objective of maximizing total flow in traffic engineering. [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: Satisfied demand of traffic engineering under zero, [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Hierarchical Placement Learning for Network Slice Provisioning

    cs.NI 2025-08 conditional novelty 4.0 of 10

    A hierarchical multi-armed bandit algorithm, HELIOS, learns cluster-then-node placement for network slice requests and reports higher acceptance with low utilization in simulations.

Reference graph

Works this paper leans on

70 extracted references · 70 canonical work pages · cited by 1 Pith paper

  1. [1]

    Contracting Wide-area Network Topologies to Solve Flow Problems Quickly

    Firas Abuzaid, Srikanth Kandula, Behnaz Arzani, Ishai Menache, Matei Zaharia, and Peter Bailis. Contracting Wide-area Network Topologies to Solve Flow Problems Quickly. In18th USENIX Symposium on Networked Systems Design and Implementation (NSDI 21), pages 175–200, 2021

    work page 2021

  2. [2]

    Allocation of Fungible Resources via a Fast, Scalable Price Discov- ery Method.Mathematical Programming Computation, 14(3):593–622, 2022

    Akshay Agrawal, Stephen Boyd, Deepak Narayanan, Fiodar Kazhamiaka, and Matei Zaharia. Allocation of Fungible Resources via a Fast, Scalable Price Discov- ery Method.Mathematical Programming Computation, 14(3):593–622, 2022

    work page 2022

  3. [3]

    A rewriting system for convex op- timization problems.Journal of Control and Decision, 5(1):42–60, 2018

    Akshay Agrawal, Robin Verschueren, Steven Diamond, and Stephen Boyd. A rewriting system for convex op- timization problems.Journal of Control and Decision, 5(1):42–60, 2018

    work page 2018

  4. [4]

    Mathematical Programming the State of the Art: Bonn

    Achim Bachem, Martin Grötschel, and Bernhard Korte. Mathematical Programming the State of the Art: Bonn

    work page

  5. [5]

    Parallelism-Optimizing Data Placement for Faster Data-Parallel Computations

    Nirvik Baruah, Peter Kraft, Fiodar Kazhamiaka, Pe- ter Bailis, and Matei Zaharia. Parallelism-Optimizing Data Placement for Faster Data-Parallel Computations. Proceedings of the VLDB Endowment, 16(4):760–771, 2022

    work page 2022

  6. [6]

    Bertsekas.Constrained optimization and Lagrange multiplier methods

    Dimitri P. Bertsekas.Constrained optimization and Lagrange multiplier methods. Academic press, 2014

    work page 2014

  7. [7]

    Branch and bound methods.Notes for EE364b, Stanford Univer- sity, 2006:07, 2007

    Stephen Boyd and Jacob Mattingley. Branch and bound methods.Notes for EE364b, Stanford Univer- sity, 2006:07, 2007

    work page 2006

  8. [8]

    Distributed Optimization and Statistical Learning via the Alternating Direction Method of Multipliers.Foundations and Trends in Ma- chine learning, 3(1):1–122, 2011

    Stephen Boyd, Neal Parikh, Eric Chu, Borja Peleato, Jonathan Eckstein, et al. Distributed Optimization and Statistical Learning via the Alternating Direction Method of Multipliers.Foundations and Trends in Ma- chine learning, 3(1):1–122, 2011

    work page 2011

Show all 70 references
  1. [9]

    A Faster Algorithm for Between- ness Centrality.Journal of mathematical sociology, 25(2):163–177, 2001

    Ulrik Brandes. A Faster Algorithm for Between- ness Centrality.Journal of mathematical sociology, 25(2):163–177, 2001

    2001

  2. [10]

    On variants of shortest-path between- ness centrality and their generic computation.Social networks, 30(2):136–145, 2008

    Ulrik Brandes. On variants of shortest-path between- ness centrality and their generic computation.Social networks, 30(2):136–145, 2008

    2008

  3. [11]

    The Direct Extension of ADMM for Multi-block Convex Minimization Problems is Not Necessarily Con- vergent.Mathematical Programming, 155(1):57–79, 2016

    Caihua Chen, Bingsheng He, Yinyu Ye, and Xiaoming Yuan. The Direct Extension of ADMM for Multi-block Convex Minimization Problems is Not Necessarily Con- vergent.Mathematical Programming, 155(1):57–79, 2016

    2016

  4. [12]

    NetHint: White-Box Networking for Multi-Tenant Data Centers

    Jingrong Chen, Hong Zhang, Wei Zhang, Liang Luo, Jeffrey Chase, Ion Stoica, and Danyang Zhuo. NetHint: White-Box Networking for Multi-Tenant Data Centers. In19th USENIX Symposium on Networked Systems De- sign and Implementation (NSDI 22), pages 1327–1343, 2022

    2022

  5. [13]

    Cohen, Yin Tat Lee, and Zhao Song

    Michael B. Cohen, Yin Tat Lee, and Zhao Song. Solving linear programs in the current matrix multiplication time. Journal of the ACM (JACM), 68(1):1–39, 2021

    2021

  6. [14]

    Workload-aware Database Monitoring and Consolidation

    Carlo Curino, Evan PC Jones, Samuel Madden, and Hari Balakrishnan. Workload-aware Database Monitoring and Consolidation. InProceedings of the 2011 ACM SIGMOD International Conference on Management of data, pages 313–324, 2011

    2011

  7. [15]

    CVXPY: A Python- embedded modeling language for convex optimization

    Steven Diamond and Stephen Boyd. CVXPY: A Python- embedded modeling language for convex optimization. Journal of Machine Learning Research, 17(83):1–5, 2016

    2016

  8. [16]

    ECOS: An SOCP solver for embedded systems

    Alexander Domahidi, Eric Chu, and Stephen Boyd. ECOS: An SOCP solver for embedded systems. In2013 European control conference (ECC), pages 3071–3076. IEEE, 2013

    2013

  9. [17]

    Eason, Xueqi He, Richard Cziva, Max Noor- mohammadpour, Srivatsan Balasubramanian, Satya- jeet Singh Ahuja, and Biao Lu

    John P. Eason, Xueqi He, Richard Cziva, Max Noor- mohammadpour, Srivatsan Balasubramanian, Satya- jeet Singh Ahuja, and Biao Lu. Hose-based Cross-layer Backbone Network Design with Benders Decomposi- tion. InProceedings of the ACM SIGCOMM 2023 Con- ference, pages 333–345, 2023

    2023

  10. [18]

    Data on Notable AI Models, 6 2024

    Epoch AI. Data on Notable AI Models, 6 2024. Ac- cessed: 2025-05-12

    2024

  11. [19]

    Machine Learning Hardware, 6 2024

    Epoch AI. Machine Learning Hardware, 6 2024. Ac- cessed: 2025-05-12

    2024

  12. [20]

    A Dual Algorithm for the Solution of Nonlinear Variational Problems via Finite Element Approximation.Computers & mathe- matics with applications, 2(1):17–40, 1976

    Daniel Gabay and Bertrand Mercier. A Dual Algorithm for the Solution of Nonlinear Variational Problems via Finite Element Approximation.Computers & mathe- matics with applications, 2(1):17–40, 1976

    1976

  13. [21]

    Roland Glowinski and Americo Marroco. Sur l’approximation, par éléments finis d’ordre un, et la ré- solution, par pénalisation-dualité d’une classe de prob- lèmes de Dirichlet non linéaires.Revue française d’automatique, informatique, recherche opérationnelle. Analyse numériqu...

    1975

  14. [22]

    Gurobi Optimizer Refer- ence Manual, 2023

    Gurobi Optimization, LLC. Gurobi Optimizer Refer- ence Manual, 2023

    2023

  15. [23]

    Hestenes

    Magnus R. Hestenes. Multiplier and Gradient Meth- ods.Journal of optimization theory and applications, 4(5):303–320, 1969

    1969

  16. [24]

    Achieving High Utilization with Software-driven WAN

    Chi-Yao Hong, Srikanth Kandula, Ratul Mahajan, Ming Zhang, Vijay Gill, Mohan Nanduri, and Roger Watten- hofer. Achieving High Utilization with Software-driven WAN. InProceedings of the ACM SIGCOMM 2013 Conference on SIGCOMM, pages 15–26, 2013

    2013

  17. [25]

    User’s Manual for CPLEX.Inter- national Business Machines Corporation, 46(53):157, 2009

    IBM ILOG CPLEX. User’s Manual for CPLEX.Inter- national Business Machines Corporation, 46(53):157, 2009

    2009

  18. [26]

    AI Benchmark: All about Deep Learning on Smartphones in 2019

    Andrey Ignatov, Radu Timofte, Andrei Kulik, Seung- soo Yang, Ke Wang, Felix Baum, Max Wu, Lirong Xu, and Luc Van Gool. AI Benchmark: All about Deep Learning on Smartphones in 2019. In2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), pages 3617–3635. I...

    2019

  19. [27]

    Patil, Joseph E

    Paras Jain, Sam Kumar, Sarah Wooders, Shishir G. Patil, Joseph E. Gonzalez, and Ion Stoica. Skyplane: Optimiz- ing Transfer Cost and Throughput Using Cloud-Aware Overlays. In20th USENIX Symposium on Networked Systems Design and Implementation, pages 1375–1389, 2023

    2023

  20. [28]

    Suhas Jayaram Subramanya, Daiyaan Arfeen, Shouxu Lin, Aurick Qiao, Zhihao Jia, and Gregory R. Ganger. Sia: Heterogeneity-aware, goodput-optimized ML- cluster scheduling. InProceedings of the 29th Sympo- sium on Operating Systems Principles, pages 642–657, 2023

    2023

  21. [29]

    Analysis of Large-Scale Multi-Tenant GPU clusters for DNN training workloads

    Myeongjae Jeon, Shivaram Venkataraman, Amar Phan- ishayee, Junjie Qian, Wencong Xiao, and Fan Yang. Analysis of Large-Scale Multi-Tenant GPU clusters for DNN training workloads. In2019 USENIX Annual Technical Conference (USENIX ATC 19), pages 947– 960, 2019

    2019

  22. [30]

    PCF: Provably Resilient Flexible Routing

    Chuan Jiang, Sanjay Rao, and Mohit Tawarmalani. PCF: Provably Resilient Flexible Routing. InProceedings of the Annual conference of the ACM Special Interest Group on Data Communication on the applications, technologies, architectures, and protocols for computer communication, ...

    2020

  23. [31]

    A New Polynomial-time Algo- rithm for Linear Programming

    Narendra Karmarkar. A New Polynomial-time Algo- rithm for Linear Programming. InProceedings of the sixteenth annual ACM symposium on Theory of comput- ing, pages 302–311, 1984

    1984

  24. [32]

    Decentralized cloud wide-area network traffic engineering with BlastShield

    Umesh Krishnaswamy, Rachee Singh, Nikolaj Bjørner, and Himanshu Raj. Decentralized cloud wide-area network traffic engineering with BlastShield. In19th USENIX Symposium on Networked Systems Design and Implementation (NSDI 22), pages 325–338, 2022

    2022

  25. [33]

    Bissonnette, Nikolaj Bjørner, Zahira Nasrin, Sonal Kothari, Prabhakar Reddy, John Abeln, Srikanth Kandula, et al

    Umesh Krishnaswamy, Rachee Singh, Paul Mattes, Paul- Andre C. Bissonnette, Nikolaj Bjørner, Zahira Nasrin, Sonal Kothari, Prabhakar Reddy, John Abeln, Srikanth Kandula, et al. OneW AN is better than two: Unifying a split W AN architecture. In20th USENIX Symposium on Networked ...

    2023

  26. [34]

    Semi-oblivious Traffic Engineering: The Road not Taken

    Praveen Kumar, Yang Yuan, Chris Yu, Nate Foster, Robert Kleinberg, Petr Lapukhov, Chiun Lin Lim, and Robert Soulé. Semi-oblivious Traffic Engineering: The Road not Taken. In15th USENIX Symposium on Net- worked Systems Design and Implementation (NSDI 18), pages 157–170, 2018

    2018

  27. [35]

    Madhyastha, and Mosharaf Chowdhury

    Fan Lai, Xiangfeng Zhu, Harsha V . Madhyastha, and Mosharaf Chowdhury. Oort: Efficient federated learning via guided participant selection. In15th USENIX Sympo- sium on Operating Systems Design and Implementation, pages 19–35, 2021

    2021

  28. [36]

    Deep Learning GPU Benchmarks, 2024

    Lambda. Deep Learning GPU Benchmarks, 2024

    2024

  29. [37]

    Le, Xiao Sun, Mosharaf Chowdhury, and Zhen- hua Liu

    Tan N. Le, Xiao Sun, Mosharaf Chowdhury, and Zhen- hua Liu. Allox: Compute Allocation in Hybrid Clusters. InProceedings of the Fifteenth European Conference on Computer Systems, pages 1–16, 2020

    2020

  30. [38]

    Shard Manager: A Generic Shard Management Framework for Geo-distributed Applications

    Sangmin Lee, Zhenhua Guo, Omer Sunercan, Jun Ying, Thawan Kooburat, Suryadeep Biswal, Jun Chen, Kun Huang, Yatpang Cheung, Yiding Zhou, et al. Shard Manager: A Generic Shard Management Framework for Geo-distributed Applications. InProceedings of the ACM SIGOPS 28th Symposium o...

    2021

  31. [39]

    Chap- man

    Sleiman Mhanna, Gregor Verbiˇc, and Archie C. Chap- man. Adaptive ADMM for Distributed AC Optimal Power Flow.IEEE Transactions on Power Systems, 34(3):2025–2035, 2018

    2025

  32. [40]

    FlexW AN: Software Hardware Co-design for Cost-Effective and Resilient Optical Backbones

    Congcong Miao, Zhizhen Zhong, Ying Zhang, Kunling He, Fangchao Li, Minggang Chen, Yiren Zhao, Xiang Li, Zekun He, Xianneng Zou, et al. FlexW AN: Software Hardware Co-design for Cost-Effective and Resilient Optical Backbones. InProceedings of the ACM SIG- COMM 2023 Conference, ...

    2023

  33. [41]

    Jordan, et al

    Philipp Moritz, Robert Nishihara, Stephanie Wang, Alexey Tumanov, Richard Liaw, Eric Liang, Melih Eli- bol, Zongheng Yang, William Paul, Michael I. Jordan, et al. Ray: A Distributed Framework for Emerging AI Applications. In13th USENIX symposium on operating systems design and...

    2018

  34. [42]

    Mind- ing the Gap between Fast Heuristics and their Optimal Counterparts

    Pooria Namyar, Behnaz Arzani, Ryan Beckett, Santiago Segarra, Himanshu Raj, and Srikanth Kandula. Mind- ing the Gap between Fast Heuristics and their Optimal Counterparts. InProceedings of the 21st ACM Workshop on Hot Topics in Networks, pages 138–144, 2022

    2022

  35. [43]

    Solving Max- Min Fair Resource Allocations Quickly on Large Graphs

    Pooria Namyar, Behnaz Arzani, Srikanth Kandula, Santi- ago Segarra, Daniel Crankshaw, Umesh Krishnaswamy, Ramesh Govindan, and Himanshu Raj. Solving Max- Min Fair Resource Allocations Quickly on Large Graphs. In21th USENIX Symposium on Networked Systems Design and Implementati...

    2024

  36. [44]

    Solving Large-Scale Granular Resource Allocation Problems Efficiently with POP

    Deepak Narayanan, Fiodar Kazhamiaka, Firas Abuzaid, Peter Kraft, Akshay Agrawal, Srikanth Kandula, Stephen Boyd, and Matei Zaharia. Solving Large-Scale Granular Resource Allocation Problems Efficiently with POP. InProceedings of the ACM SIGOPS 28th Symposium on Operating Syste...

    2021

  37. [45]

    Heterogeneity-Aware Cluster Scheduling Policies for Deep Learning Workloads

    Deepak Narayanan, Keshav Santhanam, Fiodar Kazhamiaka, Amar Phanishayee, and Matei Zaharia. Heterogeneity-Aware Cluster Scheduling Policies for Deep Learning Workloads. In14th USENIX Symposium on Operating Systems Design and Implementation, pages 481–498, 2020

    2020

  38. [46]

    Nelder and Roger Mead

    John A. Nelder and Roger Mead. A Simplex Method for Function Minimization.The computer journal, 7(4):308– 313, 1965

    1965

  39. [47]

    RAS: Continuously Optimized Region-Wide Datacen- ter Resource Allocation

    Andrew Newell, Dimitrios Skarlatos, Jingyuan Fan, Pa- van Kumar, Maxim Khutornenko, Mayank Pundir, Yirui Zhang, Mingjun Zhang, Yuanlai Liu, Linh Le, et al. RAS: Continuously Optimized Region-Wide Datacen- ter Resource Allocation. InProceedings of the ACM SIGOPS 28th Symposium ...

    2021

  40. [48]

    DOTE: Rethinking (Predictive) WAN Traffic Engineering

    Yarin Perry, Felipe Vieira Frujeri, Chaim Hoch, Srikanth Kandula, Ishai Menache, Michael Schapira, and Aviv Tamar. DOTE: Rethinking (Predictive) WAN Traffic Engineering. In20th USENIX Symposium on Networked Systems Design and Implementation (NSDI 23), pages 1557–1581, 2023

    2023

  41. [49]

    Accordion: Elastic Scalability for Database Systems Supporting Distributed Transactions.Proceedings of the VLDB Endowment, 7(12):1035–1046, 2014

    Marco Serafini, Essam Mansour, Ashraf Aboulnaga, Kenneth Salem, Taha Rafiq, and Umar Farooq Minhas. Accordion: Elastic Scalability for Database Systems Supporting Distributed Transactions.Proceedings of the VLDB Endowment, 7(12):1035–1046, 2014

    2014

  42. [50]

    TACCL: Guiding Collective Algorithm Synthesis using Com- munication Sketches

    Aashaka Shah, Vijay Chidambaram, Meghan Cowan, Saeed Maleki, Madan Musuvathi, Todd Mytkowicz, Ja- cob Nelson, Olli Saarikivi, and Rachee Singh. TACCL: Guiding Collective Algorithm Synthesis using Com- munication Sketches. In20th USENIX Symposium on Networked Systems Design and...

    2023

  43. [51]

    Cost-effective Cloud Edge Traffic En- gineering with CASCARA

    Rachee Singh, Sharad Agarwal, Matt Calder, and Paramvir Bahl. Cost-effective Cloud Edge Traffic En- gineering with CASCARA. In18th USENIX Sympo- sium on Networked Systems Design and Implementation (NSDI 21), pages 201–216, 2021

    2021

  44. [52]

    Cost-effective Capacity Provisioning in Wide Area Networks with Shoofly

    Rachee Singh, Nikolaj Bjorner, Sharon Shoham, Yawei Yin, John Arnold, and Jamie Gaudette. Cost-effective Capacity Provisioning in Wide Area Networks with Shoofly. InProceedings of the 2021 ACM SIGCOMM 2021 Conference, pages 534–546, 2021

    2021

  45. [53]

    Elmore, Ashraf Aboulnaga, Andrew Pavlo, and Michael Stonebraker

    Rebecca Taft, Essam Mansour, Marco Serafini, Jennie Duggan, Aaron J. Elmore, Ashraf Aboulnaga, Andrew Pavlo, and Michael Stonebraker. E-store: Fine-grained Elastic Partitioning for Distributed Transaction Process- ing Systems.Proceedings of the VLDB Endowment, 8(3):245–256, 2014

    2014

  46. [54]

    A Simple Effective Heuristic for Embedded Mixed-integer Quadratic Programming.In- ternational journal of control, 93(1):2–12, 2020

    Reza Takapoui, Nicholas Moehle, Stephen Boyd, and Alberto Bemporad. A Simple Effective Heuristic for Embedded Mixed-integer Quadratic Programming.In- ternational journal of control, 93(1):2–12, 2020

    2020

  47. [55]

    Learning to Route

    Asaf Valadarsky, Michael Schapira, Dafna Shahaf, and Aviv Tamar. Learning to Route. InProceedings of the 16th ACM workshop on hot topics in networks, pages 185–191, 2017

    2017

  48. [56]

    RDC: Energy- Efficient Data Center Network Congestion Relief with Topological Reconfigurability at the Edge

    Weitao Wang, Dingming Wu, Sushovan Das, Afsaneh Rahbar, Ang Chen, and TS Eugene Ng. RDC: Energy- Efficient Data Center Network Congestion Relief with Topological Reconfigurability at the Edge. In19th USENIX Symposium on Networked Systems Design and Implementation, pages 1267–1...

    2022

  49. [57]

    Global Con- vergence of ADMM in Nonconvex Nonsmooth Opti- mization.Journal of Scientific Computing, 78:29–63, 2019

    Yu Wang, Wotao Yin, and Jinshan Zeng. Global Con- vergence of ADMM in Nonconvex Nonsmooth Opti- mization.Journal of Scientific Computing, 78:29–63, 2019

    2019

  50. [58]

    Examination of WAN Traffic Characteristics in a Large-scale Data Center Network

    Zhaohua Wang, Zhenyu Li, Guangming Liu, Yunfei Chen, Qinghua Wu, and Gang Cheng. Examination of WAN Traffic Characteristics in a Large-scale Data Center Network. InProceedings of the 21st ACM Inter- net Measurement Conference, pages 1–14, 2021

    2021

  51. [59]

    Be- ware of Fragmentation: Scheduling GPU-Sharing Work- loads with Fragmentation Gradient Descent

    Qizhen Weng, Lingyun Yang, Yinghao Yu, Wei Wang, Xiaochuan Tang, Guodong Yang, and Liping Zhang. Be- ware of Fragmentation: Scheduling GPU-Sharing Work- loads with Fragmentation Gradient Descent. In2023 USENIX Annual Technical Conference, USENIX ATC ’23. USENIX Association, 2023

    2023

  52. [60]

    Lucien Werner, Adam Wierman, and Steven H. Low. Pricing Flexibility of Shiftable Demand in Electricity Markets. InProceedings of the Twelfth ACM Inter- national Conference on Future Energy Systems, pages 1–14, 2021

    2021

  53. [61]

    lp-Box ADMM: A Versatile Framework for Integer Programming.IEEE transactions on pattern analysis and machine intelli- gence, 41(7):1695–1708, 2018

    Baoyuan Wu and Bernard Ghanem. lp-Box ADMM: A Versatile Framework for Integer Programming.IEEE transactions on pattern analysis and machine intelli- gence, 41(7):1695–1708, 2018

    2018

  54. [62]

    Can’t be Late: Optimizing Spot Instance Savings Under Deadlines

    Zhanghao Wu, Wei-Lin Chiang, Ziming Mao, Zongheng Yang, Eric Friedman, Scott Shenker, and Ion Stoica. Can’t be Late: Optimizing Spot Instance Savings Under Deadlines. In21st USENIX Symposium on Networked Systems Design and Implementation (NSDI 24), pages 185–203, 2024

    2024

  55. [63]

    Gandiva: Introspective Cluster Scheduling for Deep Learning

    Wencong Xiao, Romil Bhardwaj, Ramachandran Ram- jee, Muthian Sivathanu, Nipun Kwatra, Zhenhua Han, Pratyush Patel, Xuan Peng, Hanyu Zhao, Quanlu Zhang, et al. Gandiva: Introspective Cluster Scheduling for Deep Learning. In13th USENIX Symposium on Oper- ating Systems Design and...

    2018

  56. [64]

    An Empirical Study of ADMM for Nonconvex Problems.arXiv preprint arXiv:1612.03349, 2016

    Zheng Xu, Soham De, Mario Figueiredo, Christoph Studer, and Tom Goldstein. An Empirical Study of ADMM for Nonconvex Problems.arXiv preprint arXiv:1612.03349, 2016

    2016 arXiv

  57. [65]

    Yan, Rachee Singh, Justin T

    Zhiying Xu, Francis Y . Yan, Rachee Singh, Justin T. Chiu, Alexander M. Rush, and Minlan Yu. Teal: Learning-Accelerated Optimization of W AN Traffic En- gineering. InProceedings of the ACM SIGCOMM 2023 Conference, pages 378–393, 2023

    2023

  58. [66]

    SkyPilot: An Intercloud Broker for Sky Comput- ing

    Zongheng Yang, Zhanghao Wu, Michael Luo, Wei- Lin Chiang, Romil Bhardwaj, Woosuk Kwon, Siyuan Zhuang, Frank Sifei Luan, Gautam Mittal, Scott Shenker, et al. SkyPilot: An Intercloud Broker for Sky Comput- ing. In20th USENIX Symposium on Networked Systems Design and Implementati...

    2023

  59. [67]

    Zeta: A Scalable and Robust East-West Communication Framework in Large-Scale Clouds

    Qianyu Zhang, Gongming Zhao, Hongli Xu, Zhuolong Yu, Liguang Xie, Yangming Zhao, Chunming Qiao, Ying Xiong, and Liusheng Huang. Zeta: A Scalable and Robust East-West Communication Framework in Large-Scale Clouds. In19th USENIX Symposium on Networked Systems Design and Implemen...

    2022

  60. [68]

    Cerebro: A Plat- form for Multi-Party Cryptographic Collaborative Learn- ing

    Wenting Zheng, Ryan Deng, Weikeng Chen, Raluca Ada Popa, Aurojit Panda, and Ion Stoica. Cerebro: A Plat- form for Multi-Party Cryptographic Collaborative Learn- ing. In30th USENIX Security Symposium (USENIX Security 21), pages 2723–2740, 2021

    2021

  61. [69]

    ARROW: Restoration-Aware Traffic Engineering

    Zhizhen Zhong, Manya Ghobadi, Alaa Khaddaj, Jonathan Leach, Yiting Xia, and Ying Zhang. ARROW: Restoration-Aware Traffic Engineering. InProceedings of the 2021 ACM SIGCOMM 2021 Conference, pages 560–579, 2021

    2021

  62. [70]

    PDAS: A Practical Distributed ADMM System for Large-Scale Linear Programming Problems at Alipay

    Jun Zhou, Yang Bao, Daohong Jian, and Hua Wu. PDAS: A Practical Distributed ADMM System for Large-Scale Linear Programming Problems at Alipay. InProceed- ings of the 29th ACM SIGKDD Conference on Knowl- edge Discovery and Data Mining, 2023. Appendix A Evaluation Setup of Clust...

    2023

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.