Source-linked AI summary
A Wavelength Borrowing Architecture for Optical Data Center Networks - Extended Version
Andrea Detti, Chiara Lodovisi, Silvello Betti
TL;DR
The paper addresses how optical data-center spines can accommodate unbalanced east-west traffic without the cost and power burden of electronic switching. It proposes wavelength borrowing combined with two-hop traffic engineering, controlled by a tunable borrowing degree. The architecture achieves near-optimal performance below full reconfigurability while using mature components and targeting demand that persists for seconds or longer.
Problem
East-west traffic is growing, while electronic spine switching creates cost, power, and capacity-scaling challenges and optical wavelength assignment must adapt to uneven demand.
Method
The paper combines a spine-leaf optical architecture, dynamic wavelength borrowing, two-hop traffic engineering, and a tunable borrowing degree using mature optical components.
Results
Near-optimal performance is achieved with partial rather than full reconfigurability, with B = 8 ≪32 reported as sufficient in the evaluated setting.
Takeaways & Limitations
Reconfigurability can be scaled to data-center needs instead of adopting a fully reconfigurable fabric, while the architecture targets traffic lasting seconds or longer.
Takeaways & Limitations
The paper does not consider placement engineering, which acts upstream at the traffic-source level to reduce network load.
Abstract
from arXiv · showhide
The growth of east-west traffic, along with the cost and power consumption of electronic switching, is motivating the integration of a low-power, high-rate, all-optical layer within the data center network. This paper presents a spine-leaf all-optical architecture in which the default wavelength configuration, one wavelength per source--destination leaf pair, can be reconfigured to accommodate unbalanced traffic demand: wavelengths that are unused or lightly loaded at one leaf are borrowed by another leaf with higher demand. This topology-engineering capability is combined with a traffic-engineering scheme, based on two-hop detouring, enabling the control of wavelength load while limiting the amount of detoured traffic. A key feature of the architecture is that its degree of wavelength reconfigurability is set by a single tunable parameter, the borrowing degree $B$, ranging from none to full; performance evaluation shows that near-optimal performance is achieved well below the maximum $B$, saving the complexity and cost of fully reconfigurable solutions. Furthermore, the optical fabric relies on mature, data-center-grade components, namely AWGs, AWGRs, colorless OXCs, and combiners, whose reconfiguration speed makes the architecture deployable at network tiers where traffic demand persists over seconds or longer, e.g., among groups of racks (pods). The architecture is also TDMA-transparent, a property that future work could exploit to refine the borrowing unit below a whole wavelength without changing the optical fabric.
1. INTRODUCTION
East-west traffic and the cost, power, and scaling constraints of electronic spine layers motivate all-optical data-center networks. The paper proposes a mature-component architecture whose reconfigurability and traffic engineering adapt wavelength allocation to demand while allowing complexity to scale gradually.
- East-west traffic dominates modern data-center volume, driven by distributed applications, storage replication, and distributed training and inference.
- Electronic spine scaling requires switch or transceiver replacement, additional fibers, rewiring, or extra switches when capacity must increase.
- All-optical spines reduce power consumption and remain transparent to data format and rate by carrying leaf-to-leaf traffic without intermediate opto-electronic conversion.
- The proposed architecture borrows idle wavelengths between leaves and uses two-hop detours to adapt allocation while controlling wavelength load and detoured traffic.
- Its borrowing degree provides tunable reconfigurability, avoiding the need for full reconfigurability while preserving near-optimal performance below the maximum setting.
- The design uses mature optical components and targets traffic persisting for seconds or longer, while excluding placement engineering from its scope.
A. Overview
The architecture connects pod-level leaf nodes through an SDN-managed optical spine using WDM and optional TDMA. It reallocates idle default wavelengths through borrowing, with B controlling the flexibility–hardware trade-off.
- Each pod connects to a leaf through a ToR switch, while the optical spine carries inter-leaf circuits using WDM and optional TDMA.
- Default operation assigns one wavelength per destination to each leaf, balancing optical resources across leaves.
- A donor leaf can lend an idle default wavelength to a borrower, increasing the borrower’s available bandwidth toward a destination.
- Borrowing degree B limits concurrent lending or borrowing to B −1 leaves and trades greater allocation flexibility for additional optical hardware.
- Optional TDMA reduces borrowing granularity from a full wavelength to individual time slots, enabling finer demand matching at higher hardware complexity.
B. Components and Wiring
The architecture combines modular leaf transmitters and receivers with a spine fabric built from AWGRs, combiners, and a colorless OxC. The borrowing degree B bounds wavelength-relocation resources and the number of concurrently supported donors or borrowers.
- Leaf nodes: Each leaf contains B transmission modules, each using W fixed-wavelength lasers and a W×1 AWG multiplexer.
- Leaf nodes: The default transmission module connects directly to its combiner, while B −1 borrowing modules connect through the spine OxC and can use wavelengths from distinct donors.At most B −1 donors can serve a leaf concurrently.
- Leaf nodes: An ingress SDN-controlled load balancer distributes outgoing traffic across the transmission-module buffers according to the selected traffic-engineering strategy.
- Leaf nodes: At reception, an AWG demultiplexer separates the W wavelengths, and dedicated WDM receivers forward traffic to its destination or to the next-hop leaf for detouring.
- Spine optical switching fabric: The spine fabric combines a colorless OxC, combiners, and a cyclic AWGR; combiners merge default and borrowing fibers before AWGR routing.The OxC routes borrowing fibers to combiners under SDN control, while the cyclic AWGR provides deterministic wavelength routing.
- Spine optical switching fabric: Combiner size B limits each donor to serving at most B −1 borrowers, while passive combining incurs an optical loss of 10 log10(B) dB compensated by an amplifier.The wavelength assignment keeps the wavelength sets entering each combiner disjoint, enabling a passive coupler.
- Spine optical switching fabric: Using separate BTX and BCOMB parameters would make relocation asymmetric: borrowers could use at most BTX −1 donors, while donors could serve at most BCOMB −1 borrowers.
- Spine optical switching fabric: A WSS alternative offers near-zero combining loss but requires coordinated active control, higher cost, and potentially time-slot-level reconfiguration for TDMA.
C. Leaf-to-Leaf Data Transfer
The architecture handles unbalanced leaf-to-leaf demand by borrowing underused default wavelengths and combining the added optical capacity with two-hop traffic detouring. An SDN controller coordinates borrowing, wavelength assignment, detouring, and load constraints under hardware and traffic-feasibility limits.
- Balanced configuration: Under balanced traffic, each leaf uses one default wavelength per destination, with no wavelength borrowing.Leaf 1 and leaf W reach leaf 2 using their respective default wavelengths λ2 and λ1.
- Borrowing operation: When demand is unbalanced, an underloaded donor can lend its default wavelength toward a destination to an overloaded borrower.The controller detects underutilized donor capacity and congested borrower wavelengths, then checks borrowing feasibility and benefit.
- Borrowing operation: Borrowing gives the borrower additional optical capacity while any residual donor traffic is detoured through other leaves with active paths to the destination.In the example, leaf W receives λ2 from leaf 1 and can use λ1 and λ2 to reach leaf 2.
- Traffic engineering: The controller reconfigures borrower load balancing across default and borrowed wavelengths and routes donor traffic over feasible two-hop paths when needed.If borrowing is infeasible or inconvenient, detouring alone can still reduce overload at the source leaf.
- Optimization: The formulation represents borrowing with binary variables and detouring with traffic fractions, while enforcing hardware limits, connectivity, traffic conservation, and a wavelength-load threshold.The objective minimizes total electronically detoured traffic while keeping each source–destination load below ρth.
- Practical considerations: The heuristic is intended for practical use, but its computational complexity and optimality gap are not formally analyzed.For 64 leaves and B = 16, a raw Python implementation took approximately 90 s on a 2019 i9 Intel MacBook.
5. PERFORMANCE EVALUATION
The evaluation compares static AWGR-based baselines with wavelength borrowing and two-hop water-filling under increasingly imbalanced traffic. Partial borrowing substantially reduces detouring and packet loss, achieving near-optimal performance with limited reconfigurability while scaling favorably.
- Evaluation setup: The simulator compares the borrowing heuristic with three static AWGR-only baselines using different detouring strategies.The baselines share the same static optical core, isolating the effects of traffic engineering and optical reconfiguration.
- Detouring and imbalance: Higher traffic variability requires more detouring at a fixed borrowing degree, while increasing B reduces detoured traffic.Uniform detouring reaches nearly 97% of traffic for W = 32, whereas borrowing exploits greater optical reconfiguration to reduce detouring.
- Detouring and imbalance: 2HWF alone substantially reduces detoured traffic without optical reconfiguration, but further improvement requires borrowing modules with B > 1.The B = 1 case is equivalent to AWGR-only with two-hop water-filling detouring.
- Reconfigurability: B = 8 achieves performance close to optimal in a W = 32 system, despite full reconfigurability requiring B = W.At B = 8, wavelength allocation follows the traffic pattern, and every pair remains below ρth = 0.9 with no packet loss.
- Packet loss: B > 2 ensures zero packet loss across the tested cv range, whereas B = 1 avoids loss only up to cv = 1.Uniform detouring can perform worse than AWGR-only for cv ≤1 because it approximately doubles wavelength load.
- Scale-out: B = 2 suffices for larger networks with W = 32 or 64, although the number of wavelengths must still scale linearly with the number of leaves.Larger networks provide more donor opportunities per borrowing degree, reducing the required borrowing modules per leaf.
6. RELATED WORKS
Related work spans fast, fine-grained optical fabrics near servers and slower aggregation-level fabrics connecting leaves or pods. The paper positions mature, colorless and passive technologies as practical for smoother higher-tier traffic, while contrasting them with more costly or immature alternatives.
- Technology maturity: WSS-based and optical-signal-processing architectures remain more expensive, complex, or limited in port count than passive AWGR and colorless OxC alternatives.The comparison frames mature colorless and passive components as more practical for near-term deployment.
- Deployment tiers: Optical fabrics near servers or racks face bursty traffic that can change within tens to hundreds of nanoseconds.Such deployments require packet- or sub-packet-scale reconfiguration for efficient utilization.
- Fast optical fabrics: Sirius provides sub-nanosecond end-to-end reconfiguration through tunable lasers and round-robin TDMA, but relies on uniform-topology detouring.Its passive AWGR network replaces the electrical hierarchy above the ToR.
- Fast optical fabrics: OPSquare, HFOS, ROTOS, and PULSE use packet-level optical switching approaches with substantial hardware or scalability demands.The described constraints include research-grade label processing, SOA-based switches, per-ToR WSSs, loss, and limited scalability.
- Deployment tiers: Aggregation-level fabrics connecting leaves or pods handle smoother demand evolving over milliseconds to seconds, enabling slower mature switching technologies.The cited technologies include colorless MEMS-OxCs, actively reconfigurable WSSs, and passive AWGRs.
- Aggregation-level fabrics: Jupiter demonstrates aggregation-level deployment using colorless MEMS-OxCs with traffic-engineering-driven reconfiguration on timescales of seconds to minutes.This matches the slower demand dynamics of higher-level network interconnection fabrics.
The Proposed Architecture
The proposed architecture uses mature optical components and a single borrowing degree B to tune the trade-off between reconfigurability, cost, and complexity. Combined with traffic engineering, it achieves near-optimal performance at moderate, scale-independent B without requiring full reconfigurability.
- Technology and fabric: The architecture uses passive AWGRs, colorless MEMS-based OxCs, and passive combiners, avoiding WSSs, optical signal processing, and actively reconfigurable colored components.Its fabric remains TDMA-transparent and does not require destination-leaf reconfiguration when borrowed wavelengths are activated or released.
- Reconfigurability: The borrowing degree B provides a tunable trade-off between hardware cost and wavelength-allocation flexibility.The architecture can be configured from limited to full reconfigurability by adjusting B.
- Performance and scaling: Near-optimal performance is achieved at a small B regardless of network scale, while borrowing lines per leaf can decrease as the data center grows.Larger networks provide more donor leaves and optimization opportunities for the same B.
- Operation and optimization: The architecture targets aggregation-level traffic that persists for seconds or longer and uses millisecond-scale MEMS OxC reconfiguration.A formal MILP and a practical greedy heuristic based on 2HWF jointly address wavelength assignment and traffic detouring.
A. Small Data Center
When fewer leaves than wavelengths are required, the optical fabric remains unchanged and unused default wavelengths can be borrowed by existing leaves.
- Small Data Center: The architecture assumes L equals W in its baseline description, but supports smaller deployments without changing the fabric.The unused capacity comes from the default wavelengths of absent leaves.
- Small Data Center: If the number of leaves L is smaller than the number of wavelengths W, default wavelengths assigned to absent leaves can be borrowed by existing leaves.The spine optical fabric does not need modification in this case.
B. Large Data Center
The architecture scales bandwidth through parallel optical layers and scales nodes through grouped leaves connected by multiple source–destination optical fabrics. These extensions preserve shared optical components while expanding capacity or node count.
- Bandwidth scaling: With W = 64 and S = 400 Gbit/s, the balanced configuration provides about 0.8 Pbit/s of bidirectional bisection bandwidth.The balanced bisection bandwidth is W^2S/2, where S is the per-wavelength bitrate.
- Bandwidth scaling: P parallel AWGR-routed layers increase capacity while sharing the colorless OxC, whose size does not grow with P.Each layer carries default and borrowed wavelengths, allowing bandwidth scale-out without enlarging the shared OxC.
- Node scaling: When more than W leaves are required, nodes are partitioned into G groups of at most W nodes, with G^2 distinct fabrics for ordered source–destination group pairs.Each fabric handles communication within or between a particular pair of groups.
- Node scaling: Figure 10 represents leaves as (i, j), where i identifies the leaf within group j, and central fabric (t, r) connects transmitters in group t to receivers in group r.The figure depicts a representative cross-connecting subset rather than every leaf and interconnection.
- Node scaling: Each transmitting leaf has G default fibers for group-specific traffic and B − 1 borrowing fibers for wavelengths lent by donor leaves.Borrowed wavelengths may connect leaves across arbitrary group pairs, requiring borrowing fibers to be steered to the appropriate fabric.
APPENDIX II: DERIVATION OF THE TWO-HOP ABSORBED TRAFFIC
The appendix derives the traffic absorbed by a two-hop path by limiting the detoured flow according to the more constrained hop. The resulting rule respects the target load and excludes paths whose links are already overloaded.
- Traffic absorbed by a two-hop path: The derivation treats Fθ,j,i,d as the amount of traffic absorbed by two-hop path (j, i, d) at water level θ.The same detoured traffic traverses both hops, although the induced load increments may differ.
- Per-hop limits: For each hop, the maximum absorbable traffic is computed by raising its new load to the target level θ.This provides the per-link limits used to bound the two-hop flow.
- Per-hop limits: The two-hop absorbed traffic cannot exceed either per-link limit because the same quantity traverses both hops.The more constrained hop saturates first and determines the path’s absorbed traffic.
- Target-load property: The hop attaining the minimum reaches ρnew = θ, while the other remains below θ, so the two-hop load reaches the target without exceeding it.This follows the two-hop load definition ρj,i,d = max(ρj,i, ρi,d).
- Overloaded-hop handling: If a hop already has ρold > θ, its absorbable amount is clamped to zero because an overloaded link cannot absorb negative traffic.This excludes paths that are already above the target and yields the final absorbed-traffic rule.
APPENDIX III: COMPARISON OF RELATED ARCHITECTURES
Table 2 compares all-optical data-center network architectures and identifies whether they use optical processing.
- Table 2 compares different all-optical data-center network architectures.
- “Optical processing” denotes in-network forwarding decisions made in the optical domain.
- The table uses optical processing as one criterion for comparing the architectures.