Acknowledgements: This blog was written in collaboration with Sudharsana Venkataraman and Vishnu Pavan Beeram
Introduction
Random Network Graphs (RNGs) have emerged as an intriguing alternative to traditional Clos fabrics for data centers. Rather than organizing switches into a rigid hierarchy of leaf, spine, and super-spine layers, RNGs employ a richly connected, expander-like topology that offers numerous independent paths between any pair of switches. The objective is straightforward: make network bandwidth more fungible so that traffic is less constrained by structural bottlenecks.
Amazon’s recently published work on Random Network Graphs introduces Spraypoint, a distributed routing mechanism specifically designed to exploit this path diversity. Instead of attempting to identify a single “best” path, Spraypoint spreads traffic across many possible forwarding paths, relying on the statistical properties of an expander graph to achieve high throughput.
At first glance, this raises an interesting question. If the topology already provides abundant path diversity and Spraypoint naturally exploits it, is there still a role for explicit traffic engineering?
I believe the answer is yes.
Multipath Traffic Engineering (MPTE) approaches the same problem from an entirely different perspective. Rather than assuming that a “good” topology naturally leads to good traffic placement, MPTE treats routing as an optimization problem. It explicitly computes multipath forwarding DAGs using topology, residual link capacity, and optional traffic demand information to determine how traffic should traverse the network.
Although both approaches exploit the same underlying path diversity, they optimize for fundamentally different objectives. Furthermore, as AI workloads push fabrics toward sustained high utilization, the distinction between topology-driven and optimization-driven forwarding becomes economically significant.
AI Changes the Economics
Historically, routing protocols were evaluated primarily in terms of convergence speed, forwarding simplicity, and operational scalability.
AI clusters introduce an entirely different economic model.
The cost of a poor forwarding decision is no longer measured only in packet latency. Every congested collective communication may leave thousands of expensive GPUs waiting idle. Under these conditions, even modest improvements in bandwidth utilization can translate directly into improved infrastructure efficiency.
Viewed through this lens, explicit traffic engineering becomes much easier to justify.
Spraypoint provides an elegant, scalable forwarding mechanism that effectively exploits the statistical properties of Random Graph Networks. For general-purpose traffic, it represents an attractive distributed solution requiring relatively little centralized coordination.
MPTE addresses a different operational objective.
Rather than merely discovering available paths, it attempts to maximize usable bandwidth by explicitly controlling how traffic is distributed across those paths. Residual capacity, demand volume, and optional traffic matrices all become inputs to the forwarding decision.
As AI fabrics continue operating closer to full utilization, this distinction becomes increasingly significant.
Two Different Design Philosophies
The easiest way to understand the distinction between Spraypoint and MPTE is to recognize that they solve different problems.
Spraypoint begins with the observation that expander-like topologies already possess tremendous path diversity. Instead of attempting to compute optimal routes, every router independently performs destination-oriented forwarding based upon waypoint structures relative to each destination. Packets are initially sprayed across multiple neighbors using ECMP before converging toward the destination through waypoint levels.
Example graph with Spraypoint levels for t (src: https://arxiv.org/pdf/2604.15261)
The beauty of this design lies in its simplicity. Every router independently computes forwarding information from its link-state database, requiring no centralized controller and no explicit optimization. Assuming the underlying graph has strong expansion properties, the network naturally exposes many nearly edge-disjoint paths.
MPTE asks an entirely different question.
Instead of asking how to expose path diversity, MPTE asks how to best utilize the available path diversity.
Rather than allowing every router to independently discover forwarding alternatives, MPTE computes an explicit Multipath Traffic Engineering DAG (MPTED). Candidate paths are first identified using metric slack around the shortest path. A max-flow algorithm, such as push-relabel, then computes how traffic should be distributed across those paths based upon available link capacity. The resulting forwarding DAG and weighted next-hop assignments are signaled to participating routers.
The distinction is subtle but important. Spraypoint assumes that a good topology naturally produces good traffic placement. MPTE explicitly computes good traffic placement from a good topology.
Path Diversity Versus Traffic Optimization
Both Spraypoint and MPTE benefit enormously from Random Graph Networks because the topology provides numerous alternative paths between endpoints. The difference lies in how those paths are used.
Spraypoint is fundamentally demand-oblivious. Though Spraypoint may react tactically to observed congestion, but these are reactive responses to local conditions rather than planned allocation of network resources against known demand. Every flow is treated similarly regardless of its bandwidth requirements or the current utilization of the network. The forwarding decision is based upon destination-relative waypoint structures rather than the instantaneous state of the fabric.
For general-purpose traffic, this approach works remarkably well. Expander graphs naturally distribute traffic across many independent paths, reducing the likelihood that any single cut becomes overloaded.
However, modern AI data centers increasingly operate under conditions where this assumption becomes less attractive.
Large collective communication jobs may consume hundreds of gigabits per second between specific groups of servers/GPUs. Multiple jobs may simultaneously compete for the same portions of the network. Under these conditions, knowing that multiple paths exist is only part of the problem. The more important question becomes which of those paths still have available capacity.
This is where MPTE changes the conversation.
Rather than treating every candidate path equally, MPTE can incorporate residual bandwidth into its computation. Links that are already heavily utilized contribute less available capacity during the max-flow calculation, naturally steering traffic toward less congested portions of the fabric. Even without a complete traffic matrix, MPTE continuously adapts forwarding decisions to current network conditions instead of relying solely on the statistical properties of the topology.
When a traffic matrix is available, the optimization becomes even more powerful.
Instead of computing each source-destination pair independently, MPTE can jointly optimize multiple demands, size forwarding DAGs according to expected traffic volume, avoid overloading shared bottlenecks, and even provide admission feedback to workload schedulers. Heavy AI collectives may receive broad forwarding DAGs, while smaller flows consume only the network resources they require.
In this sense, MPTE transforms path diversity from a structural property of the topology into an explicitly managed resource.
Topology Matters—Until It Doesn’t
Another interesting distinction between the two approaches is their dependence on topology.
Spraypoint was specifically designed for expander-like random graphs. Its effectiveness depends upon the assumption that spraying traffic into the network will quickly expose numerous edge-disjoint paths before converging near the destination. Under those conditions, the forwarding algorithm efficiently utilizes the natural properties of the graph.
However, real production networks are rarely ideal.
Maintenance events, hardware failures, incremental expansions, and operational constraints inevitably alter the topology over time. As expansion properties degrade, spraying traffic may increase path length without exposing meaningful additional capacity because many of the candidate paths ultimately reconverge across the same bottlenecks.
MPTE does not require the topology to be an expander or to satisfy any specific structural property. It operates on whatever topology is presented. Its effectiveness scales with the path diversity available: a richer topology yields higher max-flow, but the algorithm adapts gracefully to constrained topologies.
Whether the network is a Clos fabric, Dragonfly, partial mesh, random graph, or a degraded expander after multiple failures, MPTE computes forwarding DAGs using the topology that actually exists. If failures reduce available capacity across a particular cut, the optimization naturally accounts for the reduced residual bandwidth during the max-flow computation.
This does not imply that topology no longer matters. On the contrary, richly connected Random Graph Networks provide MPTE with many more candidate paths from which to build forwarding DAGs. In many respects, RNG and MPTE complement one another exceptionally well.
The difference is simply that Spraypoint depends upon the topology being an expander, while MPTE simply takes advantage of whatever topology is available.
Conclusion
Spraypoint and MPTE should not be viewed as competing implementations of the same idea. They embody two fundamentally different philosophies for operating highly connected data-center fabrics.
Spraypoint trusts the topology. It assumes that a well-designed expander graph naturally provides sufficient path diversity and that distributed spraying will efficiently exploit that diversity.
MPTE trusts optimization. It assumes that path diversity represents an opportunity rather than a complete solution and that explicit computation can place traffic more effectively than statistical load balancing alone.
Both approaches leverage the same rich connectivity provided by Random Graph Networks, yet they pursue different operational goals.
As data-center networks evolve to support increasingly communication-intensive AI workloads, the discussion may shift away from whether topology or traffic engineering is more important. Instead, the more interesting question may become how the two can work together—combining the resilience and simplicity of topology-driven forwarding with the precision and adaptability of utilization-aware traffic engineering.
References
[RNG] RNG: Flat Datacenter Networks at Scale https://arxiv.org/pdf/2604.15261
[MPTE] Multipath Traffic Engineering https://datatracker.ietf.org/doc/draft-kompella-teas-mpte/
Leave a Reply