pith. sign in

arxiv: 2605.28205 · v1 · pith:UIYDUODWnew · submitted 2026-05-27 · 💻 cs.DC

Resource Allocation in HyperX Networks

Alejandro Cano , Crist\'obal Camarero , Carmen Mart\'inez , Ram\'on Beivide This is my paper

Pith reviewed 2026-06-29 09:55 UTC · model grok-4.3

classification 💻 cs.DC
keywords resource allocationHyperX networksHPC systemsnetwork topologyperformance evaluationdiagonal allocationpartition bandwidth
0
0 comments X

The pith

The non-convex diagonal allocation strategy outperforms traditional resource allocations in HyperX networks by improving partition bandwidth and switch locality.

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

The paper formalizes resource allocation in HyperX networks as linear, geometric, and stochastic functions and analyzes their effects on network properties such as dilation, convexity, and partition bandwidth. Experiments with synthetic traffic patterns and application kernels show that partition bandwidth and switch locality determine how well allocations avoid interference under various routing algorithms. The diagonal strategy, despite lacking convexity, produces the best results across most tested scenarios. If correct, this means HPC systems built on HyperX can achieve lower communication overhead simply by changing how processes are mapped to nodes, without altering the underlying topology or routing.

Core claim

Resource allocation strategies for HyperX networks are categorized into linear, geometric, and stochastic functions. Theoretical characterization of their topological properties shows that the diagonal allocation, which is not convex, yields higher partition bandwidth and better switch locality than convex alternatives. Exhaustive evaluation under synthetic traffic and application kernels confirms that these properties reduce interference and improve performance for multiple routing algorithms.

What carries the argument

The diagonal allocation strategy, a geometric mapping function that places processes along network diagonals to increase locality.

If this is right

  • Partition bandwidth is a decisive factor that mitigates interferences more effectively than convexity alone.
  • Switch locality influences performance outcomes under different routing algorithms.
  • A set of lessons learned can guide the design of resource allocation policies for HyperX-based HPC systems.
  • Geometric strategies can be superior to linear or stochastic ones when evaluated on realistic kernels.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same diagonal mapping idea could be tested on other richly connected low-diameter topologies to check whether non-convexity is broadly beneficial.
  • If the diagonal strategy scales with network size, it may reduce the need for custom routing optimizations in future HyperX deployments.
  • Production schedulers could incorporate a quick diagonal pre-mapping step before launching jobs to gain immediate locality benefits.

Load-bearing premise

The synthetic traffic patterns and application communication kernels used in the evaluation represent the workloads that arise in actual HyperX deployments.

What would settle it

A direct measurement of communication time or throughput on a real HyperX system running production applications, comparing the diagonal strategy against linear and convex geometric baselines under the same routing algorithm.

Figures

Figures reproduced from arXiv: 2605.28205 by Alejandro Cano, Carmen Mart\'inez, Crist\'obal Camarero, Ram\'on Beivide.

Figure 1
Figure 1. Figure 1: Three topologies employed in HPC. From left to right: a 54-server, 6-radix, 3-level Fat-tree, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: shows a 2D HyperX with 5 switches per dimension and 125 endpoints. Observe that every row and column is a complete graph and, conse￾quently, the network diameter is 2. 0,0 0,1 0,2 0,3 0,4 1,0 1,1 1,2 1,3 1,4 2,0 2,1 2,2 2,3 2,4 3,0 3,1 3,2 3,3 3,4 4,0 4,1 4,2 4,3 4,4 [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Minimal (solid) and non-minimal routing paths (dotted) in Omni-WAR routing in a 5 × 5 HyperX. Links are omitted for clarity. 2.2 Routing To deal with different traffic patterns, the net￾work must be equipped with an adaptive routing mechanism. A well-established routing for Hy￾perX topologies is Omni-WAR [22]. A closely re￾lated mechanism that uses the same set of routes is DAL [2]. The algorithm for Omni-… view at source ↗
Figure 4
Figure 4. Figure 4: Scalability of HyperX and other topolo￾gies. specifically allowing more than one deroute in the same dimension. Thus, it allows each packet to per￾form up to q+m hops, where m is the non-minimal hops permitted, and q is the number of dimensions in the HyperX. For some adverse traffic patterns it is necessary that m ≥ q to achieve acceptable per￾formance, and m = q is always sufficient; this is assumed henc… view at source ↗
Figure 5
Figure 5. Figure 5: Inter-group communication from reduce scatter a la Rabenseifner of 64 processes with 8 groups of 8 processes each. Each group is mapped to the servers attached to one switch of a 8 × 8 HX. The 8 selected switches lie in a diagonal of the topology. the host graph. When considering multiple appli￾cations, the embedding can be considered jointly, by taking as guest graph the disjoint union of all applications… view at source ↗
Figure 6
Figure 6. Figure 6: Some linear and tile based partitions for [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Makespan of 64-process kernels with Omni-WAR routing. The light-colored bars show the time running isolated and the dark part shows the extra time taken due to the interference. seen, Random and Full Spread are the worst strate￾gies as the ones most troubled by interference and HoL blocking. Then Rectangles, L-shapes, Rows and Random Switch Selection follow, with similar performance. Finally, Diagonal is t… view at source ↗
Figure 9
Figure 9. Figure 9: Makespan for load escalation of 64-processes kernels with Omni-WAR routing, allocating in [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Makespan for different 64-process kernels. Background traffic is a random permutation. The [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Makespan of 64-process kernels provid￾ing a different VC set to the background traffic. The dark part shows the extra time taken due to the interference over running isolated. not reported here for brevity, as they yielded simi￾lar conclusions to the other cases. 7.3 Fabric partitioning Figures 11 and 12 show interference analysis re￾sults with more virtual channels, for both basic traffic and kernel traf… view at source ↗
Figure 12
Figure 12. Figure 12: Makespan of 64-process kernels providing a different VC set to the background traffic. The [PITH_FULL_IMAGE:figures/full_fig_p017_12.png] view at source ↗
read the original abstract

As high-performance computing systems scale in size and complexity, efficient resource management is essential to minimize communication overhead. The HyperX is a richly connected, low-diameter network that offers a scalable and cost-effective alternative to traditional topologies. However, resource allocation in HyperX remains underexplored, and strategies designed for networks like Torus, Fat-tree, or Dragonfly do not directly transfer. In this work, we propose and formalize several resource allocation strategies for HyperX networks, categorized into linear, geometric, and stochastic functions. We characterize these strategies theoretically by analyzing their topological properties, including dilation, convexity, and partition bandwidth.Furthermore, we conduct an exhaustive experimental evaluation using synthetic traffic and application communication kernels to assess the impact of these strategies on performance under different routing algorithms. Our results indicate that partition bandwidth and switch locality are decisive factors in mitigating interferences. Notably, the Diagonal allocation strategy, which is not convex, consistently outperforms traditional approaches in most scenarios. Finally, we provide a set of lessons learned to guide the implementation of resource allocation policies in HPC systems based on HyperX networks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper proposes and formalizes resource allocation strategies for HyperX networks, categorized into linear, geometric, and stochastic functions. It characterizes these strategies theoretically via topological properties including dilation, convexity, and partition bandwidth. An exhaustive experimental evaluation using synthetic traffic patterns and application communication kernels under different routing algorithms finds that partition bandwidth and switch locality are decisive in mitigating interference, and that the non-convex Diagonal strategy consistently outperforms traditional approaches in most scenarios. Lessons learned for HPC implementations are provided.

Significance. If the empirical ranking holds, the work fills a gap in HyperX resource allocation by showing that convexity is not required for strong performance and by identifying partition bandwidth as a key factor. The combination of theoretical analysis and evaluation across routing algorithms could inform allocation policies in large-scale systems. Credit is due for the exhaustive experimental design and the explicit categorization of strategies.

major comments (1)
  1. [Experimental Evaluation] Experimental Evaluation section: The central claim that the Diagonal allocation strategy 'consistently outperforms traditional approaches in most scenarios' (abstract) is demonstrated exclusively on synthetic traffic patterns and application kernels. No argument, mapping, or sensitivity analysis is provided to establish that these patterns produce interference behaviors representative of production HyperX deployments; this directly affects whether the reported superiority and the decisiveness of partition bandwidth/switch locality generalize.
minor comments (2)
  1. [Abstract] Abstract: Reports experimental outcomes but provides no quantitative metrics, error bars, number of runs, or exclusion criteria, reducing the ability to assess result robustness from the summary alone.
  2. Notation for the three categories (linear, geometric, stochastic) and the specific Diagonal strategy is introduced without early formal definitions or pseudocode, which would aid readability before the theoretical analysis.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback and for acknowledging the exhaustive experimental design. We address the single major comment point by point below.

read point-by-point responses

  1. Referee: [Experimental Evaluation] Experimental Evaluation section: The central claim that the Diagonal allocation strategy 'consistently outperforms traditional approaches in most scenarios' (abstract) is demonstrated exclusively on synthetic traffic patterns and application kernels. No argument, mapping, or sensitivity analysis is provided to establish that these patterns produce interference behaviors representative of production HyperX deployments; this directly affects whether the reported superiority and the decisiveness of partition bandwidth/switch locality generalize.

    Authors: We agree that the manuscript provides no explicit argument, mapping, or sensitivity analysis connecting the chosen synthetic patterns and kernels to interference behaviors in actual production HyperX deployments. The evaluation relies on standard HPC traffic models (uniform random, bit-reversal, and application kernels) that are widely used to isolate partition-bandwidth and locality effects. To address the concern, we will revise the Experimental Evaluation section by adding a dedicated paragraph that (1) cites prior literature validating these patterns for high-radix network studies and (2) reports a limited sensitivity sweep over traffic intensity and locality parameters. This addition will clarify the intended scope of generalization while leaving the reported rankings and the identified importance of partition bandwidth unchanged. revision: partial

Circularity Check

0 steps flagged

No circularity detected; claims rest on independent theoretical analysis and experiments

full rationale

The paper proposes and formalizes allocation strategies, then characterizes them via topological properties (dilation, convexity, partition bandwidth) and evaluates performance experimentally on synthetic traffic and kernels. No equations, fitted parameters, or self-citations are shown that reduce any reported result or prediction to the inputs by construction. The Diagonal strategy outperformance is presented as an empirical observation from the evaluation suite, not a fitted or self-defined quantity.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No free parameters, axioms, or invented entities are identifiable from the abstract alone.

pith-pipeline@v0.9.1-grok · 5725 in / 987 out tokens · 26264 ms · 2026-06-29T09:55:14.881523+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

28 extracted references · 1 canonical work pages · 1 internal anchor

  1. [1]

    The intel programmable and integrated unified memory architecture graph analytics processor.IEEE Micro, 43(5):78–87, 2023

    Sriram Aananthakrishnan, Shamsul Abe- din, Vincent Cav´ e, Fabio Checconi, Kristof Du Bois, Stijn Eyerman, Joshua B Fryman, Wim Heirman, Jason Howard, Ibrahim Hur, et al. The intel programmable and integrated unified memory architecture graph analytics processor.IEEE Micro, 43(5):78–87, 2023

    2023

  2. [2]

    Schreiber

    Jung Ho Ahn, Nathan Binkert, Al Davis, Moray McLaren, and Robert S. Schreiber. Hy- perX: Topology, routing, and packaging of ef- ficient large-scale networks. InProceedings of the Conference on High Performance Comput- ing Networking, Storage and Analysis, SC ’09, pages 1–11, New York, NY, USA, 2009. ACM

    2009

  3. [3]

    Aleliunas and A.L

    R. Aleliunas and A.L. Rosenberg. On embed- ding rectangular grids in square grids.IEEE Transactions on Computers, 31(09):907–913, September 1982

    1982

  4. [4]

    Revisiting the outer-weakly convex domination number in graph products

    Bijo S Anand, Jonecis A Dayap, Leomarich F Casinillo, Karen Luz P Yap, et al. Re- visiting the outer-weakly convex domination number in graph products.arXiv preprint arXiv:2501.15524, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  5. [5]

    Bhuyan and Dharma P

    Laxmi N. Bhuyan and Dharma P. Agrawal. Generalized hypercube and hyperbus struc- tures for a computer network.Computers, IEEE Transactions on, C-33(4):323–333, April 1984

    1984

  6. [6]

    Cam- bridge university press, 1993

    Norman Biggs.Algebraic graph theory. Cam- bridge university press, 1993

    1993

  7. [7]

    Achiev- ing high-performance fault-tolerant routing in HyperX interconnection networks

    Crist´ obal Camarero, Alejandro Cano, Car- men Mart´ ınez, and Ram´ on Beivide. Achiev- ing high-performance fault-tolerant routing in HyperX interconnection networks. InSC24- W: Workshops of the International Confer- ence for High Performance Computing, Net- working, Storage and Analysis, pages 472–483, 2024

    2024

  8. [8]

    Polarized routing: an effi- cient and versatile algorithm for large direct networks

    Crist´ obal Camarero, Carmen Mart´ ınez, and Ram´ on Beivide. Polarized routing: an effi- cient and versatile algorithm for large direct networks. In2021 IEEE Symposium on High- Performance Interconnects (HOTI), pages 52– 59, 2021

    2021

  9. [9]

    The CAMINOS interconnec- tion networks simulator.Journal of Parallel and Distributed Computing, 204:105136, 2025

    Crist´ obal Camarero, Daniel Postigo, and Pablo Fuentes. The CAMINOS interconnec- tion networks simulator.Journal of Parallel and Distributed Computing, 204:105136, 2025

    2025

  10. [10]

    CAMINOS.https://crates.io/crates/ caminos

  11. [11]

    Exploiting geometric parti- tioning in task mapping for parallel computers

    Mehmet Deveci, Sivasankaran Rajamanickam, Vitus J Leung, Kevin Pedretti, Stephen L Olivier, David P Bunde, Umit V Cataly¨ urek, and Karen Devine. Exploiting geometric parti- tioning in task mapping for parallel computers. In2014 IEEE 28th international parallel and distributed processing symposium, pages 27–36. IEEE, 2014

    2014

  12. [12]

    Ivanov, Yuki Tsushima, Tomoya Yuki, Aki- hiro Nomura, Shin’ichi Miura, Nic McDonald, Dennis L

    Jens Domke, Satoshi Matsuoka, Ivan R. Ivanov, Yuki Tsushima, Tomoya Yuki, Aki- hiro Nomura, Shin’ichi Miura, Nic McDonald, Dennis L. Floyd, and Nicolas Dub´ e. HyperX topology: First at-scale implementation and comparison to the fat-tree. InProceedings of the International Conference for High Perfor- mance Computing, Networking, Storage and Analysis, SC ’...

    2019

  13. [13]

    Cray Cascade: a scal- able HPC system based on a dragonfly net- work

    Greg Faanes, Abdulla Bataineh, Duncan Roweth, Tom Court, Edwin Froese, Bob Alver- son, Tim Johnson, Joe Kopnick, Mike Higgins, and James Reinhard. Cray Cascade: a scal- able HPC system based on a dragonfly net- work. InProceedings of the International Con- ference on High Performance Computing, Net- working, Storage and Analysis, SC ’12, pages 1–9, Los Al...

    2012

  14. [14]

    Overview of the Blue Gene/L system archi- tecture.IBM Journal of research and devel- opment, 49(2.3):195–212, 2005

    Alan Gara, Matthias A Blumrich, Dong Chen, GL-T Chiu, Paul Coteus, Mark E Gi- ampapa, Ruud A Haring, Philip Heidelberger, Dirk Hoenicke, Gerard V Kopcsay, et al. Overview of the Blue Gene/L system archi- tecture.IBM Journal of research and devel- opment, 49(2.3):195–212, 2005

    2005

  15. [15]

    On the automor- phism group of a composite graph.Studia Sci

    F Harary and EM Palmer. On the automor- phism group of a composite graph.Studia Sci. Math. Hungar, 3:439–441, 1968

    1968

  16. [16]

    Quiet neighborhoods: Key to protect job performance predictabil- ity

    Ana Jokanovic, Jose Carlos Sancho, German Rodriguez, Alejandro Lucero, Cyriel Minken- berg, and Jesus Labarta. Quiet neighborhoods: Key to protect job performance predictabil- ity. In2015 IEEE International Parallel and Distributed Processing Symposium, pages 449– 459, 2015

    2015

  17. [17]

    TPU v4: An optically re- configurable supercomputer for machine learn- ing with hardware support for embeddings

    Norm Jouppi, George Kurian, Sheng Li, Pe- ter Ma, Rahul Nagarajan, Lifeng Nai, Nishant Patil, Suvinay Subramanian, Andy Swing, 19 Brian Towles, et al. TPU v4: An optically re- configurable supercomputer for machine learn- ing with hardware support for embeddings. In Proceedings of the 50th annual international symposium on computer architecture, pages 1– 14, 2023

    2023

  18. [18]

    Dally, and Dennis Abts

    John Kim, William J. Dally, and Dennis Abts. Flattened butterfly: a cost-efficient topology for high-radix networks. InProceedings of the 34th annual international symposium on Com- puter architecture, ISCA ’07, pages 126–137, New York, NY, USA, 2007. ACM

    2007

  19. [19]

    Dally, Steve Scott, and Dennis Abts

    John Kim, William J. Dally, Steve Scott, and Dennis Abts. Technology-driven, highly- scalable dragonfly topology. InProceedings of the 35th Annual International Symposium on Computer Architecture, pages 77–88. IEEE Computer Society, 2008

    2008

  20. [20]

    LaForge, Kirk F

    Laurence E. LaForge, Kirk F. Korver, and M. Sami Fadali. What designers of bus and network architectures should know about hy- percubes.Computers, IEEE Transactions on, 52(4):525–544, April 2003

    2003

  21. [21]

    Fugaku and A64FX: the first exascale supercomputer and its innovative Arm CPU

    Satoshi Matsuoka. Fugaku and A64FX: the first exascale supercomputer and its innovative Arm CPU. In2021 Symposium on VLSI Cir- cuits, pages 1–3. IEEE, 2021

    2021

  22. [22]

    Practical and efficient incremental adaptive routing for Hy- perX networks

    Nic McDonald, Mikhail Isaev, Adriana Flo- res, Al Davis, and John Kim. Practical and efficient incremental adaptive routing for Hy- perX networks. InProceedings of the Interna- tional Conference for High Performance Com- puting, Networking, Storage and Analysis, SC ’19, New York, NY, USA, 2019. Association for Computing Machinery

    2019

  23. [23]

    Interval-regular graphs.Discrete Mathematics, 41(3):253–269, 1982

    Henry Martyn Mulder. Interval-regular graphs.Discrete Mathematics, 41(3):253–269, 1982

    1982

  24. [24]

    Optimization of collec- tive reduction operations

    Rolf Rabenseifner. Optimization of collec- tive reduction operations. InComputational Science-ICCS 2004: 4th International Confer- ence, Krak´ ow, Poland, June 6-9, 2004, Pro- ceedings, Part I 4, pages 1–9. Springer, 2004

    2004

  25. [25]

    Dragonfly+: Low cost topol- ogy for scaling datacenters

    Alexander Shpiner, Zachy Haramaty, Saar Eliad, Vladimir Zdornov, Barak Gafni, and Eitan Zahavi. Dragonfly+: Low cost topol- ogy for scaling datacenters. In2017 IEEE 3rd International Workshop on High-Performance Interconnection Networks in the Exascale and Big-Data Era (HiPINEB), pages 1–8. IEEE, 2017

    2017

  26. [26]

    TOP500 list of su- percomputer sites, November 2025

    Erich Strohmaier, Jack Dongarra, Horst Si- mon, and Martin Meuer. TOP500 list of su- percomputer sites, November 2025

    2025

  27. [27]

    Hao Yu, I-Hsin Chung, and Jos´ e E. Moreira. Topology mapping for Blue Gene/L supercom- puter. InSC 2006 Conference, Proceedings of the ACM/IEEE, November 2006

    2006

  28. [28]

    Level-spread: A new job alloca- tion policy for dragonfly networks

    Yijia Zhang, Ozan Tuncer, Fulya Kaplan, Katzalin Olcoz, Vitus J Leung, and Ayse K Coskun. Level-spread: A new job alloca- tion policy for dragonfly networks. In2018 IEEE International Parallel and Distributed Processing Symposium (IPDPS), pages 1123–