Search

EP-4740402-A1 - NETWORK ARCHITECTURE WITH HARMONIC CONNECTIONS

EP4740402A1EP 4740402 A1EP4740402 A1EP 4740402A1EP-4740402-A1

Abstract

A computer network organized in a logical grid having rows and columns can include network nodes coupled according to harmonics. Each network node can be coupled to network nodes of the same row using a set of horizontal strands according to a set of horizontal harmonics. Each of the horizontal harmonics specifies a node distance along the row between adjacent connection points on the corresponding horizontal strand. Each network node can also be coupled to network nodes of the same column using a set of vertical strands according to a set of vertical harmonics. Each of the vertical harmonics specifies a node distance along the column between adjacent connection points on the corresponding vertical strand.

Inventors

  • BERNARDI, GIACOMO
  • MAHAJAN, RATUL
  • KUMAR, SAURABH

Assignees

  • Amazon Technologies, Inc.

Dates

Publication Date
20260513
Application Date
20240625

Claims (1)

  1. AMZ2P274WO_P82487-WO01 CLAIMS WHAT IS CLAIMED IS: 1. A computer network, comprising: a logical grid having rows and columns of network nodes, wherein each network node is coupled to network nodes of a same row using a set of horizontal strands according to a set of horizontal harmonics, each of the horizontal harmonics specifying a node distance along the row between adjacent connection points on a corresponding horizontal strand; and wherein each network node is coupled to network nodes of a same column using a set of vertical strands according to a set of vertical harmonics, each of the vertical harmonics specifying a node distance along the column between adjacent connection points on a corresponding vertical strand. 2. The computer network of claim 1, wherein each link between connection points on each strand in the set of horizontal strands and the set of vertical strands has a dynamically adjustable bandwidth capacity. 3. The computer network of claim 2, wherein each strand has a maximum bandwidth capacity represented as a number of channels available on the strand, and a bandwidth capacity of each link along the strand is adjusted by reallocating the number of channels available on the strand. 4. The computer network of claim 1, wherein each link between connection points on each strand in the set of horizontal strands and the set of vertical strands has a fixed bandwidth capacity. 5. The computer network of any of claims 1 to 4, wherein each strand in the set of horizontal strands and the set of vertical strands connects up to a maximum allowable number of network nodes per strand. 6. The computer network of claim 5, wherein the maximum allowable number of network nodes per strand is based on a number of channels supported by the strand. 7. The computer network of claim 6, wherein the number of channels supported by the strand is based on a number of optical wavelengths supported by the strand. AMZ2P274WO_P82487-WO01 8. The computer network of any of claims 1 to 7, wherein the set of horizontal strands connected to a network node has a same number of strands as the set of vertical strands connected to the network node. 9. The computer network of any of claims 1 to 8, wherein each horizontal harmonic in the set of horizontal harmonics is different than other horizontal harmonics in the set of horizontal harmonics, or each vertical harmonic in the set of vertical harmonics is different than other vertical harmonics in the set of vertical harmonics. 10. The computer network of any of claims 1 to 9, wherein the set of horizontal harmonics includes at least one prime number that is larger than two, or the set of vertical harmonics includes at least one prime number that is larger than two. 11. The computer network of any of claims 1 to 9, wherein the set of horizontal strands includes a horizontal strand that spans at least a full length of the row, and the set of vertical strands includes a vertical strand that spans at least a full length of the column. 12. A computer-implemented method for routing network traffic between a source node and a destination node in a network fabric containing network nodes interconnected by strands according to harmonics that each specifies a node distance between adjacent connection points on a corresponding strand, the method comprising: distributing the network traffic from the source node to a set of one-hop neighbors of the source node; and for each one-hop neighbor of the source node in the set of one-hop neighbors of the source node: setting the one-hop neighbor of the source node as a first waypoint; identifying a one-hop neighbor of the destination node that is mapped to the one-hop neighbor of the source node; setting the identified one-hop neighbor of the destination node as a second waypoint; and routing the distributed traffic from the first waypoint to the second waypoint via the network fabric. 13. The computer-implemented method of claim 12, wherein the distributed network traffic is routed using equal cost multipath routing (ECMP) for traffic AMZ2P274WO_P82487-WO01 demand between the source node and the destination node being below a threshold, and weighted cost multipath routing (WCMP) for the traffic demand being at or above the threshold. 14. The computer-implemented method of any of claims 12 to 13, wherein routing the distributed traffic is performed by an optimizer that determines an amount of traffic to carry on each path, the optimizer receiving a set of inputs including: a traffic demand matrix containing traffic information between each pair of network nodes in the network fabric; connection topology including the harmonics of the network fabric; and valid paths for each pair of network nodes in the network fabric. 15. The computer-implemented method of claim 14, wherein the set of inputs to the optimizer further includes a maximum number of channels supported per strand and bandwidth capacity per channel, and wherein the optimizer further determines channel allocation on links between connection points on each strand.

Description

AMZ2P274WO_P82487-WO01 NETWORK ARCHITECTURE WITH HARMONIC CONNECTIONS INCORPORATION BY REFERENCE [0001] A Request Form is filed concurrently with this specification as part of the present application. Each application that the present application claims benefit of or priority to as identified in the concurrently filed Request Form is incorporated herein by reference in its entirety and for all purposes. BACKGROUND [0002] Server computers such as those supporting cloud computing services are maintained in facilities commonly referred to as data centers. A small data center may occupy a room or a floor of a building, while a large data center may occupy several floors or an entire building. A typical data center may house thousands of servers that communicate with each other via a network. The computing workload demanded from data centers have increased dramatically to serve computationally intensive applications such as large machine learning models. As such, data centers are expanding in size and numbers to meet the increase in workload demands. BRIEF DESCRIPTION OF THE DRAWINGS [0003] Various embodiments in accordance with the present disclosure will be described with reference to the drawings, in which: [0004] FIG.1 illustrates an example of a strand connecting network nodes (e.g., routing devices), according to certain aspects of the disclosure; [0005] FIG.2A illustrates an example of a strand with reconfigurable channels connecting network nodes, according to certain aspects of the disclosure; [0006] FIG.2B illustrates an example of reconfiguring a channel of the strand of FIG.2A, according to certain aspects of the disclosure; [0007] FIG.3 illustrates an example of harmonic connections between network nodes, according to certain aspects of the disclosure; [0008] FIG.4A illustrates an example of network nodes in a data center, according to certain aspects of the disclosure; [0009] FIG.4B illustrates an example of harmonic connections between network nodes in a data center, according to certain aspects of the disclosure; AMZ2P274WO_P82487-WO01 [0010] FIG.5 illustrates an example of routing traffic between a source and a destination, according to certain aspects of the disclosure; [0011] FIG.6 illustrates a first example implementation of a strand, according to certain aspects of the disclosure; [0012] FIG.7 illustrates a second example implementation of a strand, according to certain aspects of the disclosure; [0013] FIG.8 illustrates a third example implementation of a strand, according to certain aspects of the disclosure; [0014] FIG.9 illustrates a fourth example implementation of a strand, according to certain aspects of the disclosure; [0015] FIG.10 illustrates a fifth example implementation of a strand, according to certain aspects of the disclosure; [0016] FIG.11 illustrates a flow diagram of an example of a process for routing traffic in a network fabric, according to certain aspects of the disclosure; [0017] FIG.12 illustrates a flow diagram of an example of a process for performing communications in a network fabric, according to certain aspects of the disclosure; and [0018] FIG.13 illustrates an example of a network device, according to certain aspects of the disclosure. DETAILED DESCRIPTION [0019] To support the expansion of data centers, the network infrastructure interconnecting the data center has to scale with the number of server computers. A multilayer network fabric such as a Fat Tree topology or other variations of Clos topologies have been popular due to their flexibility and non-blocking nature. However, such topologies may not scale linearly with the number of server racks due to the addition of spine switches and intermediate switching layers between Top-of-Rack switches and spine switches. The excess equipment also increases power consumption. This makes scaling Clos topologies beyond a mega data center expensive in terms of the need for additional network devices and overall power consumption. [0020] The techniques disclosed herein provide a network architecture that uses a single connection layer to interconnect network nodes in a computer network. For example, a computer network implemented in a data center may include multiple network nodes organized as a logical grid. Each network node can be associated with a server rack, and each logical column of network nodes may correspond to an aisle or other grouping of server racks. Each network node can be implemented, for example, using a routing device coupled AMZ2P274WO_P82487-WO01 to the servers in the server rack. The network nodes can be interconnected with each other using strands (e.g., a small optical network) implemented with multipoint optical technologies that provide optical paths between the network nodes. [0021] For example, each network node can be coupled to network nodes in the same column (e.g., on the same aisle) using a set of vertical strands, and coupled to network nodes in the same row using a set of ho