WO2013165390A1 - Configurable clos network - Google Patents

Configurable clos network Download PDF

Info

Publication number
WO2013165390A1
WO2013165390A1 PCT/US2012/035918 US2012035918W WO2013165390A1 WO 2013165390 A1 WO2013165390 A1 WO 2013165390A1 US 2012035918 W US2012035918 W US 2012035918W WO 2013165390 A1 WO2013165390 A1 WO 2013165390A1
Authority
WO
WIPO (PCT)
Prior art keywords
leaf
leafs
spine
spines
ports
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/US2012/035918
Other languages
French (fr)
Inventor
Michael Michael Schlansker
Jose Jose Renato G. Santos
Michael Michael Renne Ty Tan
Guodong Guodong Zhang
Shih-Yuan Shih-Yuan Wang
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hewlett Packard Development Co LP
Original Assignee
Hewlett Packard Development Co LP
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hewlett Packard Development Co LP filed Critical Hewlett Packard Development Co LP
Priority to CN201280074442.XA priority Critical patent/CN104380679A/en
Priority to PCT/US2012/035918 priority patent/WO2013165390A1/en
Priority to US14/398,579 priority patent/US9584373B2/en
Priority to EP12876066.7A priority patent/EP2845358A4/en
Publication of WO2013165390A1 publication Critical patent/WO2013165390A1/en
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04LTRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L49/00Packet switching elements
    • H04L49/65Re-configuration of fast packet switches
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04LTRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/08Configuration management of networks or network elements
    • H04L41/0803Configuration setting
    • H04L41/0813Configuration setting characterised by the conditions triggering a change of settings
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04LTRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L49/00Packet switching elements
    • H04L49/15Interconnection of switching modules
    • H04L49/1515Non-blocking multistage, e.g. Clos
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04QSELECTING
    • H04Q11/00Selecting arrangements for multiplex systems
    • H04Q11/0001Selecting arrangements for multiplex systems using optical switching
    • H04Q11/0005Switch and router aspects
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04QSELECTING
    • H04Q11/00Selecting arrangements for multiplex systems
    • H04Q11/0001Selecting arrangements for multiplex systems using optical switching
    • H04Q11/0062Network aspects
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04QSELECTING
    • H04Q11/00Selecting arrangements for multiplex systems
    • H04Q11/0001Selecting arrangements for multiplex systems using optical switching
    • H04Q11/0005Switch and router aspects
    • H04Q2011/0052Interconnection of switches
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04QSELECTING
    • H04Q11/00Selecting arrangements for multiplex systems
    • H04Q11/0001Selecting arrangements for multiplex systems using optical switching
    • H04Q11/0062Network aspects
    • H04Q2011/009Topology aspects
    • H04Q2011/0096Tree

Definitions

  • Clos architectural features may be used to build large scale networks; i.e., a Clos network.
  • a Clos network uses a multistage switching circuit. Clos networks are useful when the physical switching needs exceed the capacity of the largest feasible single crossbar switch.
  • a further advantage of Clos networks is that the required number of crosspoints (which make up each crossbar switch) can be much fewer than were the entire switching system implemented with one large crossbar switch.
  • Figures 1 A - 1 D are schematic representations of aspects of a Clos network
  • Figure 2 is a perspective view of an embodiment of a section of a configurable Clos network
  • Figure 3 illustrates a control system embodiment for the configurable Clos network of Figure 2;
  • Figures 4A - 4D illustrate a programming embodiment for the configurable Clos network of Figure 2;
  • Figures 5 - 9 illustrate programming embodiments for the configurable Clos network of Figure 2
  • Figure 10 illustrates an embodiment of a programmable switch used with the configurable Clos network of Figure 2
  • Figure 1 1 is a flow chart illustrating an example of a method for configuring a Clos network. Detailed Description
  • a large Clos network may require the addition of more network equipment including more compute racks or additional network switches.
  • Traditional methods require substantial rewiring as new equipment is added.
  • Disclosed herein is a configurable Clos network that allows expansion or contraction of compute and networking equipment using circuit switch reconfiguration and without rewiring existing network cables.
  • the herein disclosed configurable Clos network uses small-scale, programmable and bidirectional circuit switches to connect spine and leaf cards.
  • the switches are optical switches.
  • the switches are electrical switches. Because they are much smaller than current Clos switches, the herein disclosed switches are inexpensive and easy to fabricate, install, and program, thereby providing significant advantages over larger Clos switches.
  • a leaf, leaf switch, or leaf card refers to a connections to hosts.
  • a host may be a computer system with a network interface controller (NIC) that connects to the leaf switch port.
  • NIC network interface controller
  • a storage device also has a NIC that connects to a leaf and allows the storage device to communicate with a host.
  • a spine (or fabric), spine switch, or spine card interconnects leaf switches together and allows communications between leaf switches.
  • One specific use of the configurable Clos network is in a large scale datacenter, and if the datacenter is large enough, more than one such Clos network may be deployed.
  • Datacenters are traditionally built using separately optimized host racks and core switches.
  • the configurable Clos network can optimize the entire datacenter, thereby decreasing installation and maintenance costs and increasing flexibility for future expansion of the datacenter.
  • datacenter networks are, ideally, customized for specific customer needs including number of hosts, access bandwidth, QoS, availability and other factors.
  • Datacenters commonly require multiple networks, such as networks for: primary and failover, production and research, control and data, LAN and SAN, or interactive and batch. These multiple network requirements may necessitate complex planning and on-site customization as network components are purchased and cabled for specialized needs.
  • Customized networks are difficult to rewire with evolving scale and needs. While logical topologies are complex and configuration dependent, physical connectivity ideally should remain simple and independent of configuration. For example, cabinets may be connected using low cost future proof cables with easy installation and replacement.
  • Clos networks are also used to build large port count core switches using smaller port count switch ASICs.
  • the disclosed configurable Clos network may be used to allow the modular addition of leaf cards to provide additional network ports and spine cards to provide additional network bandwidth in such core switches.
  • the configurable Clos network addresses the design issues in a large- scale datacenter by using circuit switching.
  • the configurable Clos network provides configurable interconnect topologies to solve at least two key problems.
  • First, non-configurable interconnects lose performance when a subset of leaf cabinets and spine cards are installed. For example, when half of leafs and spines are installed on a Clos network, half of leaf ports connect to missing spines, and half of spine ports connect to missing leafs.
  • circuit switching enables partitioning of hardware implementing redundant networks for fault tolerance as well as independent networks for heterogeneous traffic segregation. Fault tolerance requires careful topology design to assign ports to fault zones as the number of cabinets and spine cards is varied.
  • Independent networks can be used, for example, to segregate lossless SAN traffic from lossy LAN traffic, or to overcome current ASIC limitations and lack of management software needed for converged fabric networks.
  • Naive circuit switch systems use a full crossbar for flexibility in connecting top-of-rack (TOR) uplinks to spines. Such solutions are expensive, however.
  • the herein disclosed configurable Clos network uses multiple small-scale optical switches (for example, 48 12-port switches to interconnect 576 TOR switch ports).
  • Figures 1 A - 1 D are schematic representations of aspects of a hypothetical Clos network.
  • Figure 1 A is a schematic representation of the connections among 8 leaf cards and 8 spine cards in a hypothetical Clos network.
  • connection illustrated is between eight, 8-port, spine cards (the columns) and eight, 8-port leaf, cards (the rows).
  • Each "X" in Figure 1 A represents a connection between the spines and the leafs, and there are 64 connections. As can be seen in Figure 1 A, all eight ports of the spine cards and all eight ports of the leaf cards are used.
  • Figure 1 B is a schematic representation of connections between 4 leaf cards and 4 spine cards in the hypothetical Clos network where the same 8-port spines and leaf cards of Figure 1 B are connected.
  • the connections are represented by an "X" and the unused ports by a square.
  • Figure 1 B only 4 ports per card are used, and 4 ports of the spine cards and 4 ports of the leaf cards are unused. These unused connections waste hardware resources and result in lowered bandwidth per connection compared to the bandwidth per connection when all ports are connected.
  • Figure 1 C illustrates a 6 x 6 connection where 3/4 of the available ports on each of the cards are used.
  • Figure 1 D illustrates a 2 x 2 connection where 1/4 of the ports on each of the cards are used.
  • the configurable Clos network addresses the problems illustrated in the connections of Figures 1 B - 1 D.
  • the disclosed configurable Clos network provides for enhanced flexibility in connecting Clos components (i.e., spine and leaf cards) over the connections shown in the hypothetical Clos network of Figures 1 A - 1 D.
  • the configurable Clos network provides for connecting each port of each leaf card to each spine card no matter how many spine cards and leaf cards are used in the network.
  • the configurable Clos network uses a small-scale, programmable optical switch arrangement in a switch fabric to bi- directionally connect the spine cards and the leaf cards.
  • the disclosed switch fabric allows each unused port in each leaf and each unused port in each spine shown in Figures 1 B - 1 D to be connected to a corresponding spine and leaf, respectively.
  • the spine and leaf card configurations disclosed below apply to varying numbers of leaf and spine cards.
  • the configurable Clos network allows for the even distribution of leaf traffic onto spine cards for each of these configurations.
  • spine cards may be provisioned across separate networks of arbitrary bandwidth as needed by a customer application.
  • Leaf and spine card numbers need not correspond. For example, a full complement of 24 leaf ASICs can be load balanced across only 8 out of a full complement of 12 spines with bandwidth added as needed in a cost proportional manner. Partial hardware configurations often result in irregular networks. For example, if 12 uplinks from one leaf are evenly distributed across 5 spines, the leaf has 3 connections to 2 spines and 2 connections to 3 spines.
  • Figure 2 is a perspective view of an embodiment of a section of a configurable Clos network in which small-scale, programmable switches are used for bi-directional signaling or communications between spine cards and leaf cards.
  • the section of the configurable Clos network is structured so that a multitude of leaf cards are arranged with each leaf card coupled to each of a multitude of interconnects, and each of the interconnects in turn is coupled to each of a multitude of spines.
  • section 100 of configurable Clos network 10 includes spine cards 1 10 that are connected to leaf connections 150 by a switch fabric 120 that includes substrates 130 on which are formed small-scale, programmable circuit switches 140.
  • the circuit switches 140 are electrical switches.
  • the circuit switches 140 are optical switches.
  • Figure 1 0 illustrates an embodiment of such an optical switch.
  • the section 1 00 shown in Figure 2 may be repeated to form a complete central cabinet structure for the configurable Clos network 1 0. However, for ease of illustration, only one section 100 is shown.
  • the spine cards 1 10 may be placed in a spine rack and the leaf cards 154 in separate leaf racks.
  • Leaf card ASICs may be connected to host computers or other devices that communicate through the configurable Clos network. Illustration of these structures and devices is not necessary to explain the operation and function of the configurable Clos network, and so they are omitted from Figure 2.
  • each leaf connection is coupled to a corresponding leaf card 154 with 12 bidirectional optical fibers 152.
  • Each leaf card 154 has formed thereon an ASIC 156.
  • Each of the 12 substrates 130 has an signal path through the connection and through one of the bi-directional fibers 152 to a distinct leaf ASIC port.
  • the section 100 includes twelve of each of the cards, substrates, connections and switches, and fewer are shown for ease of illustration.
  • the section 100 is not limited to a 12 X 12 matrix, and more or fewer cards, substrates, connections and switches may be used.
  • the section 100 of the configurable Clos network 10 the number of spine cards 1 10 and leaf connections 150 and leaf cards 154 do not have to be equal, as will be explained with respect to Figures 5 - 9.
  • the section 100 may be constructed with 1 2 leaf connections 150 and 12 leaf cards 154, and 12 spine cards 1 10, but more than 12 substrates 130 and circuit switches 140.
  • the section 100 may be constructed with 24 such substrates 130 and circuit switches 140.
  • the section 100 may include 12 spine cards with a subsection of 12 leaf cards, 12 leaf connectors, and 12 interconnects 130 with circuit switches 140 replicated any number of times (i.e., the 12 spine cards 1 1 0 each connects to many more than 12 interconnects 130).
  • the spine cards 1 1 0, leaf connections 150, and leaf cards 154 are parallel to each other, and the connection between the spine cards 1 10 and leaf connections 150 is by way of substrates 130 with their associated small-scale, programmable circuit switches 140, and with the substrates 130 positioned orthogonal to the cards 1 10 and connections 150.
  • the section 100 includes twelve substrates 130, each of which has formed thereon a 1 2 X 12 small-scale, programmable optical switch 140 that provides bi-directional signaling between the spine cards 1 10 and the leaf connections 150.
  • the section 100 may include twice as many switches 140 and substrates 130 as there are spine cards 1 10 and leaf cards 154.
  • the number of substrates 130 is chosen to match at least the number of ports on a card edge.
  • the substrates 130 are plastic, silicon, or other suitable material on which are formed optical pathways or fibers to optical circuit switches 140.
  • the substrates 130 are printed circuit boards with electrical pathways or wires to electrical circuit switches 140 formed on the printed circuit boards.
  • the spine cards 1 10 have formed thereon ASICs 1 12, and the leaf connections 1 50 are connected to the leaf cards 154 (which may include ASICs 156) by way of cables 152.
  • the cables 152 may be high-bandwidth optical fibers or electrical cables. The bandwidth may be 40 GB per second or more.
  • the ASICs 1 12 and 1 56 may be Ethernet ASICs.
  • Configuration of the configurable Clos network may be controlled by operation of the switches formed on substrates positioned between the leaf connections and the spine cards.
  • Figure 3 illustrates an control system embodiment for the configurable Clos network of Figure 2.
  • the circuit switches 140 may be controlled or operated by controller 160, which executes programming code stored in data store 162 to operate the switches 140 to open or close a signal path in the switch fabric 1 20.
  • the controller executes programming to activate or deactivate electrostatic devices 1 64 to open and close optical signal paths in the switch fabric 120.
  • the optical switches may be implemented with moveable mirrors (see Figure 10), and the controller 160 may control the electrostatic devices 164 to position the moveable mirror.
  • the programming code executed by the controller 160 may cause the controller 160 to determine the number of installed leaf and spine cards and thus the number of available leaf and spine ports. The controller 160 then may execute machine instructions of the programming code to align a maximum number of bi-directional signal paths between the leafs and spines that are connected to the switch fabric 120 of the configurable Clos network. Alternately, the controller 160 may execute machine instructions to remove a leaf or spine that has a failed component from connection in the switch fabric 120.
  • the circuit switches may have 1 2 bidirectional 20 Gb/s data ports. Two 12 X 12 circuit switch cards are used to switch 40Gb/s lanes.
  • the vertical stack of 48 circuit switches provides 1 1 .5 terabits of configurable bandwidth.
  • the configurable Clos network as implemented in a datacenter reduces cost and increases flexibility when compared to a conventional (i.e., non-configurable) datacenter.
  • a first savings results from the small-scale optical switches, which are much less expensive than large fullcrossbar optical switch. .
  • a second savings results from improved modular expandability. Without configuration capability, full use of potential bandwidth requires a full complement of spines even when the number of host cabinets and leafs is reduced. With configuration, the number of spines may be reduced in proportion to the number of host cabinets.
  • Figures 4A - 4D illustrate a programming embodiment for the configurable Clos network of Figure 2.
  • the example programming shown in Figures 4A - 4D is based on a configurable Clos network having four, 4-port, spine cards and four, 4-port, leaf cards. Four cards with four ports each are used for ease of illustration.
  • Figures 4A - 4D illustrate a top-down, simplified schematic view of a connection segment of the configurable Clos network.
  • the spine cards and leaf cards are numbered 0 - 3.
  • the spine cards and leaf cards are connected by a substrate having formed thereon a programmable switch ASIC.
  • Figure 10 illustrates an example of a programmable switch ASIC that may be programmed according to Figures 4A - 4D.
  • Figure 4A illustrates the programmed connections at the 0th layer of a segment of a configurable Clos network.
  • the leaf cards are designated 0 - 3, the spine cards 0 - 3, and the interconnect substrate and switch ASIC is designated as 0.
  • these designations are purely arbitrary. For simplicity of illustration, some details of the connection of the interconnect substrate and switch ASIC are omitted; however, Figure 4A is sufficient to illustrate programming of the 0th switch ASIC.
  • port 0 of the 0th leaf card is bi-directionally coupled to port 0 of the 0th spine card, with the dashed lines within the boundaries of the interconnect substrate and switch ASIC representing the signal paths between the leaf cards and the spine cards.
  • the Oth ports of the leaf cards 1 , 2, and 3 are coupled, respectively, to the Oth ports of the spine cards 1 , 2, and 3.
  • Figure 4B illustrates an example of programming of the 1 st switch ASIC.
  • the 1 st port of the Oth leaf card is coupled to the 1 st port of the 1 st spine card
  • the 1 st port of the 1 st leaf card is coupled to the 1 st port of the 2nd spine card
  • the 1 st port of the 2nd leaf card is coupled to the 1 st port of the 3rd spine card
  • the 1 st port of the 3rd leaf card is coupled to the 1 st port of the Oth spine card.
  • Figure 4C illustrates an example of programming of the 2nd switch ASIC. As can be seen, the programmed signal paths between the 2nd ports of the leaf and spine cards are shifted by one compared to the programmed signal paths of Figure 4B.
  • Figure 4D illustrates an example of programming of the 3rd, and last, switch ASIC. As can be seen, the programmed signal paths between the 3rd ports of the leaf and spine cards are shifted by one compared to the programmed signal paths of Figure 4C.
  • Figures 4A - 4D illustrate an example of switch ASIC programming in which each port of the leaf cards is coupled at least once to each spine card, and each port of the spine cards is coupled at least once to each leaf card.
  • This signal path shifting may be accomplished using the example mechanisms shown in Figure 10.
  • the signal path shifting may provide a somewhat uniform distribution of signals between leaf and spine cards.
  • FIGS 5 - 9 illustrate programming embodiments for the configurable Clos network of Figure 2.
  • Each spine card has 12 ports, and each leaf card has 12 ports.
  • 12 spine cards and 12 leaf cards are installed are all available ports coupled, and all available hardware ports are used.
  • the spine cards 1 1 0s are represented by the columns, and the substrates 130/switches 140 are represented by the rows.
  • the switches 140 which are represented by rows 0 - 1 1 , and labeled as substrate.
  • each crossover between a column and a row represents a connection, and each such connection is given a number, from 0 to 1 1 .
  • row 0, column 0 has a number 0 registered, meaning the first port of the first connector 150 (and by extension the first leaf card 154) connects to the first port of the first spine card 1 10.
  • the 0th port of the 0th connector 150 (and by extension, the 0th leaf card) connects to the Oth port of the Oth spine card (and vice versa, since the connection paths are bidirectional).
  • the programming illustrated for the Oth switch 140 provides for straight-through programming. Following row 0 (i.e., the first row) row 1 has the port shifted by one, such that spine card 0 (the first spine card) is the output connection to connector 1 1 (and leaf card 1 1 (the twelfth leaf card)). The process of shifting connections by one with each succeeding row proceeds until all the optical switch positions are programmed for all rows.
  • the numbers 0 - 1 1 represent the programming of the Clos switch.
  • each port of each spine card connects to each leaf card at least once, and no spine card port is connected more than once to a specific leaf card. Conversely, each port of each leaf card connects to each spine card at least once, and no leaf card port is connected more than once to a specific spine card. Moreover, the programming provides for uniformly distributed signaling with all hardware fully utilized and no unutilized spine or leaf ports.
  • Figure 6 illustrates a programming scheme when only six spine cards and six leaf cards are installed.
  • the programmable switch ensures that all available ports are in fact used, even though the number (6 each) of spines and leafs is less than the maximum number (1 2) of spines and leafs.
  • spine card 0 connects to leaf card 0 twice (the 0 at position A and the 0 at position B).
  • each spine card has exactly two connections to each of the leaf cards.
  • all the hardware ports of each spine card and each leaf card are used.
  • the available bandwidth over the two connections is higher than if only one connection is made.
  • Figure 7 illustrates an alternate programming scheme that may be used when the configurable Clos network is configured with eight spine cards and eight leaf cards.
  • the 12 inputs or uplinks available on each leaf card now are distributed over 8 spine cards.
  • row (switch 140(0)) 0 is programmed to 0 - 7
  • row 1 (switch 140(1 )) is programmed to 7 - 6, and so on.
  • each leaf card has four single and four double connections to the spine cards.
  • the left most, or 0th, leaf card is connected twice to the 0th, 7th, 6th, and 5th spine cards, and once to the 4th, 3rd, 2nd, and 1 st spine cards That is, spine card 0, for example, is connected twice input ports 0, 7, 6, and 5, and once at ports 4, 3, 2, and 1 .
  • spine card 0 for example, is connected twice input ports 0, 7, 6, and 5, and once at ports 4, 3, 2, and 1 .
  • Figure 8 illustrates an alternate programming scheme that may be implemented when the configurable Clos network is configured with 1 2 leaves and 1 0 spines.
  • row 0 is programmed 0 - 9
  • row 1 is programmed 9, 0, 1 , ... 8
  • row 2 is programmed 8, 9, 0, 1 , 2 ... 7, and so on.
  • each leaf has 8 single and 2 double connections to the spines.
  • Figure 9 illustrates the programming scheme that may be implemented when four leaf cards and four spine cards are installed, but without the four spine cards being installed in fill order. That is, the four leaf cards occupy positions 0, 2, 3, and 5, and positions 1 , 4, and 6 - 1 1 are empty of leaf cards.
  • leaf row 0 is programmed 0, 2, 3, and 5; leaf row 1 is programmed 5, 0, 2, and 3, and so on.
  • each leaf has 4 triple connections to the spines.
  • the example of Figure 9 shows connectors (leaf cards) occupying the same location numbers as the spines, although such an arrangement between connectors and spines is not necessary.
  • FIG 10 illustrates an embodiment of a programmable circuit switch used with the configurable Clos network of Figure 2.
  • small-scale programmable optical switch 140 is implemented as an ASIC on substrate 130.
  • the switch 140 uses moveable, two-sided mirrors 142 that are positioned to route signals between leaf outputs and spine inputs, and spine outputs and leaf inputs, by changing the signal path 90 degrees.
  • the mirrors 142 are angled 45 degrees relative to a path between the outputs and the inputs, and are moveable along this inclined position.
  • the mirrors are positioned in one position using a positive electrostatic force, and in the other position by a negative electrostatic force.
  • Figure 1 0 shows the connections for two leaf cards and two spine cards leaf cards (L1 and L2 and spine cards S2 and S3).
  • Mirrors 142(2) and 142(3) are positioned to route the signals from the respective cards.
  • a signal arriving at L1 0 (leaf out) is diverted by mirror 142(2) to S2
  • a signal arriving at S2 0 (spine out) is diverted by mirror 142(2) to L1
  • any output from a leaf (that is, from any LN 0 ) can be routed or directed to any spine input (that is, to any SN
  • the embodiment shown in Figure 10 is one example of a programmable circuit switch, and other circuit switch may be used to route signals between the leafs and spines.
  • a switch similar to that shown in Figure 10 can use fixed, two-sided mirrors that are "turned on” to divert a signal path from straight line to 90 degrees, rather than using moveable, two-sided mirrors.
  • two substrates, each with one-way switch ASICs could be used.
  • Figure 1 1 is a flow chart illustrating an example of a method for configuring a Clos network.
  • method 200 begins at block 205 when switch controller 1 60 determines a number of leafs and a number of spines connected to the switch fabric.
  • the controller 160 determines a number of available leaf and spine ports.
  • the controller 160 generates control signals to position the mirrors 142 to provide form a maximum number of leaf-spine optical connections. Method 200 then ends.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Data Exchanges In Wide-Area Networks (AREA)

Description

CONFIGURABLE CLOS NETWORK
Background
Clos architectural features may be used to build large scale networks; i.e., a Clos network. A Clos network uses a multistage switching circuit. Clos networks are useful when the physical switching needs exceed the capacity of the largest feasible single crossbar switch. A further advantage of Clos networks is that the required number of crosspoints (which make up each crossbar switch) can be much fewer than were the entire switching system implemented with one large crossbar switch. Description of the Drawings
The detailed description will refer to the following drawings in which like numerals refer to like items, and in which;
Figures 1 A - 1 D are schematic representations of aspects of a Clos network; Figure 2 is a perspective view of an embodiment of a section of a configurable Clos network;
Figure 3 illustrates a control system embodiment for the configurable Clos network of Figure 2;
Figures 4A - 4D illustrate a programming embodiment for the configurable Clos network of Figure 2;
Figures 5 - 9 illustrate programming embodiments for the configurable Clos network of Figure 2; Figure 10 illustrates an embodiment of a programmable switch used with the configurable Clos network of Figure 2; and
Figure 1 1 is a flow chart illustrating an example of a method for configuring a Clos network. Detailed Description
After it is deployed, a large Clos network may require the addition of more network equipment including more compute racks or additional network switches. Traditional methods require substantial rewiring as new equipment is added. Disclosed herein is a configurable Clos network that allows expansion or contraction of compute and networking equipment using circuit switch reconfiguration and without rewiring existing network cables. The herein disclosed configurable Clos network uses small-scale, programmable and bidirectional circuit switches to connect spine and leaf cards. In an example, the switches are optical switches. In another example, the switches are electrical switches. Because they are much smaller than current Clos switches, the herein disclosed switches are inexpensive and easy to fabricate, install, and program, thereby providing significant advantages over larger Clos switches. Furthermore, because, as is disclosed below, the configurable Clos network uses many of these small-scale switches, programming the switches so that available spine and leaf ports are used is facilitated. The herein disclosed configurable Clos network is more efficient than existing Clos networks at least because it may make full use of leaf and spine connections, as will be described below. As used herein, a leaf, leaf switch, or leaf card, refers to a connections to hosts. A host may be a computer system with a network interface controller (NIC) that connects to the leaf switch port. A storage device also has a NIC that connects to a leaf and allows the storage device to communicate with a host. A spine (or fabric), spine switch, or spine card, interconnects leaf switches together and allows communications between leaf switches. One specific use of the configurable Clos network is in a large scale datacenter, and if the datacenter is large enough, more than one such Clos network may be deployed. Datacenters are traditionally built using separately optimized host racks and core switches. The configurable Clos network can optimize the entire datacenter, thereby decreasing installation and maintenance costs and increasing flexibility for future expansion of the datacenter. Moreover, datacenter networks are, ideally, customized for specific customer needs including number of hosts, access bandwidth, QoS, availability and other factors. Datacenters commonly require multiple networks, such as networks for: primary and failover, production and research, control and data, LAN and SAN, or interactive and batch. These multiple network requirements may necessitate complex planning and on-site customization as network components are purchased and cabled for specialized needs. Customized networks are difficult to rewire with evolving scale and needs. While logical topologies are complex and configuration dependent, physical connectivity ideally should remain simple and independent of configuration. For example, cabinets may be connected using low cost future proof cables with easy installation and replacement.
Clos networks are also used to build large port count core switches using smaller port count switch ASICs. The disclosed configurable Clos network may be used to allow the modular addition of leaf cards to provide additional network ports and spine cards to provide additional network bandwidth in such core switches.
The configurable Clos network addresses the design issues in a large- scale datacenter by using circuit switching. The configurable Clos network provides configurable interconnect topologies to solve at least two key problems. First, non-configurable interconnects lose performance when a subset of leaf cabinets and spine cards are installed. For example, when half of leafs and spines are installed on a Clos network, half of leaf ports connect to missing spines, and half of spine ports connect to missing leafs. Second, circuit switching enables partitioning of hardware implementing redundant networks for fault tolerance as well as independent networks for heterogeneous traffic segregation. Fault tolerance requires careful topology design to assign ports to fault zones as the number of cabinets and spine cards is varied. Independent networks can be used, for example, to segregate lossless SAN traffic from lossy LAN traffic, or to overcome current ASIC limitations and lack of management software needed for converged fabric networks. Naive circuit switch systems use a full crossbar for flexibility in connecting top-of-rack (TOR) uplinks to spines. Such solutions are expensive, however. To reduce cost, the herein disclosed configurable Clos network uses multiple small-scale optical switches (for example, 48 12-port switches to interconnect 576 TOR switch ports). Figures 1 A - 1 D are schematic representations of aspects of a hypothetical Clos network. Figure 1 A is a schematic representation of the connections among 8 leaf cards and 8 spine cards in a hypothetical Clos network. The connection illustrated is between eight, 8-port, spine cards (the columns) and eight, 8-port leaf, cards (the rows). Each "X" in Figure 1 A represents a connection between the spines and the leafs, and there are 64 connections. As can be seen in Figure 1 A, all eight ports of the spine cards and all eight ports of the leaf cards are used.
Figure 1 B is a schematic representation of connections between 4 leaf cards and 4 spine cards in the hypothetical Clos network where the same 8-port spines and leaf cards of Figure 1 B are connected. The connections are represented by an "X" and the unused ports by a square. As can be seen in Figure 1 B, only 4 ports per card are used, and 4 ports of the spine cards and 4 ports of the leaf cards are unused. These unused connections waste hardware resources and result in lowered bandwidth per connection compared to the bandwidth per connection when all ports are connected.
Figure 1 C illustrates a 6 x 6 connection where 3/4 of the available ports on each of the cards are used. Figure 1 D illustrates a 2 x 2 connection where 1/4 of the ports on each of the cards are used. Thus, the connections shown in Figures 1 B - 1 D are inefficient in that available hardware is going unused, and result in lower bandwidth per connection. The configurable Clos network addresses the problems illustrated in the connections of Figures 1 B - 1 D. In addition, the disclosed configurable Clos network provides for enhanced flexibility in connecting Clos components (i.e., spine and leaf cards) over the connections shown in the hypothetical Clos network of Figures 1 A - 1 D.
More specifically, the configurable Clos network provides for connecting each port of each leaf card to each spine card no matter how many spine cards and leaf cards are used in the network. The configurable Clos network uses a small-scale, programmable optical switch arrangement in a switch fabric to bi- directionally connect the spine cards and the leaf cards. The disclosed switch fabric allows each unused port in each leaf and each unused port in each spine shown in Figures 1 B - 1 D to be connected to a corresponding spine and leaf, respectively.
The spine and leaf card configurations disclosed below apply to varying numbers of leaf and spine cards. Using a number of the disclosed small-scale switches, the configurable Clos network allows for the even distribution of leaf traffic onto spine cards for each of these configurations. Similarly, spine cards may be provisioned across separate networks of arbitrary bandwidth as needed by a customer application. Leaf and spine card numbers need not correspond. For example, a full complement of 24 leaf ASICs can be load balanced across only 8 out of a full complement of 12 spines with bandwidth added as needed in a cost proportional manner. Partial hardware configurations often result in irregular networks. For example, if 12 uplinks from one leaf are evenly distributed across 5 spines, the leaf has 3 connections to 2 spines and 2 connections to 3 spines.
Figure 2 is a perspective view of an embodiment of a section of a configurable Clos network in which small-scale, programmable switches are used for bi-directional signaling or communications between spine cards and leaf cards. In an embodiment, the section of the configurable Clos network is structured so that a multitude of leaf cards are arranged with each leaf card coupled to each of a multitude of interconnects, and each of the interconnects in turn is coupled to each of a multitude of spines.
In Figure 2, section 100 of configurable Clos network 10 includes spine cards 1 10 that are connected to leaf connections 150 by a switch fabric 120 that includes substrates 130 on which are formed small-scale, programmable circuit switches 140. In an embodiment, the circuit switches 140 are electrical switches. In another embodiment, the circuit switches 140 are optical switches. Figure 1 0 illustrates an embodiment of such an optical switch. The section 1 00 shown in Figure 2 may be repeated to form a complete central cabinet structure for the configurable Clos network 1 0. However, for ease of illustration, only one section 100 is shown. Furthermore, the spine cards 1 10 may be placed in a spine rack and the leaf cards 154 in separate leaf racks. Leaf card ASICs may be connected to host computers or other devices that communicate through the configurable Clos network. Illustration of these structures and devices is not necessary to explain the operation and function of the configurable Clos network, and so they are omitted from Figure 2.
In Figure 2, four spine cards 1 10, three substrates 1 30, three circuit switches 140, and four leaf connections 150 are illustrated. Each leaf connection is coupled to a corresponding leaf card 154 with 12 bidirectional optical fibers 152. Each leaf card 154 has formed thereon an ASIC 156. Each of the 12 substrates 130 has an signal path through the connection and through one of the bi-directional fibers 152 to a distinct leaf ASIC port. However, as will be discussed with respect to Figures 5 - 9, the section 100 includes twelve of each of the cards, substrates, connections and switches, and fewer are shown for ease of illustration. Furthermore, the section 100 is not limited to a 12 X 12 matrix, and more or fewer cards, substrates, connections and switches may be used. Still further, in the section 100 of the configurable Clos network 10, the number of spine cards 1 10 and leaf connections 150 and leaf cards 154 do not have to be equal, as will be explained with respect to Figures 5 - 9. Yet further, the section 100 may be constructed with 1 2 leaf connections 150 and 12 leaf cards 154, and 12 spine cards 1 10, but more than 12 substrates 130 and circuit switches 140. For example, the section 100 may be constructed with 24 such substrates 130 and circuit switches 140. Even further, in an embodiment, the section 100 may include 12 spine cards with a subsection of 12 leaf cards, 12 leaf connectors, and 12 interconnects 130 with circuit switches 140 replicated any number of times (i.e., the 12 spine cards 1 1 0 each connects to many more than 12 interconnects 130).
As can be seen in Figure 2, the spine cards 1 1 0, leaf connections 150, and leaf cards 154 are parallel to each other, and the connection between the spine cards 1 10 and leaf connections 150 is by way of substrates 130 with their associated small-scale, programmable circuit switches 140, and with the substrates 130 positioned orthogonal to the cards 1 10 and connections 150.
In an embodiment, because there are twelve each of the cards 1 10 and connections 1 50, the section 100 includes twelve substrates 130, each of which has formed thereon a 1 2 X 12 small-scale, programmable optical switch 140 that provides bi-directional signaling between the spine cards 1 10 and the leaf connections 150. However, as noted above, the section 100 may include twice as many switches 140 and substrates 130 as there are spine cards 1 10 and leaf cards 154. In general, the number of substrates 130 is chosen to match at least the number of ports on a card edge. In an embodiment, the substrates 130 are plastic, silicon, or other suitable material on which are formed optical pathways or fibers to optical circuit switches 140. Alternately, the substrates 130 are printed circuit boards with electrical pathways or wires to electrical circuit switches 140 formed on the printed circuit boards. Finally, the spine cards 1 10 have formed thereon ASICs 1 12, and the leaf connections 1 50 are connected to the leaf cards 154 (which may include ASICs 156) by way of cables 152. The cables 152 may be high-bandwidth optical fibers or electrical cables. The bandwidth may be 40 GB per second or more. The ASICs 1 12 and 1 56 may be Ethernet ASICs. Configuration of the configurable Clos network may be controlled by operation of the switches formed on substrates positioned between the leaf connections and the spine cards. Figure 3 illustrates an control system embodiment for the configurable Clos network of Figure 2. The circuit switches 140 may be controlled or operated by controller 160, which executes programming code stored in data store 162 to operate the switches 140 to open or close a signal path in the switch fabric 1 20. In an alternative, the controller executes programming to activate or deactivate electrostatic devices 1 64 to open and close optical signal paths in the switch fabric 120. The optical switches may be implemented with moveable mirrors (see Figure 10), and the controller 160 may control the electrostatic devices 164 to position the moveable mirror.
The programming code executed by the controller 160 may cause the controller 160 to determine the number of installed leaf and spine cards and thus the number of available leaf and spine ports. The controller 160 then may execute machine instructions of the programming code to align a maximum number of bi-directional signal paths between the leafs and spines that are connected to the switch fabric 120 of the configurable Clos network. Alternately, the controller 160 may execute machine instructions to remove a leaf or spine that has a failed component from connection in the switch fabric 120.
The circuit switches may have 1 2 bidirectional 20 Gb/s data ports. Two 12 X 12 circuit switch cards are used to switch 40Gb/s lanes. The vertical stack of 48 circuit switches provides 1 1 .5 terabits of configurable bandwidth. The configurable Clos network as implemented in a datacenter reduces cost and increases flexibility when compared to a conventional (i.e., non-configurable) datacenter. A first savings results from the small-scale optical switches, which are much less expensive than large fullcrossbar optical switch. . A second savings results from improved modular expandability. Without configuration capability, full use of potential bandwidth requires a full complement of spines even when the number of host cabinets and leafs is reduced. With configuration, the number of spines may be reduced in proportion to the number of host cabinets. Oversubscribed configurations are supported by further reducing the number of spines in proportion to the oversubscription factor. A third savings results from the spine cabinet's ability to deploy components across multiple networks. For example, while traditional networks require separate spare spines for LAN and SAN fail-over, with the herein disclosed configurable Clos network, one spare spine may be shared between LAN and SAN failover functions. Finally, important savings result from reduced deployment cost as many of the complexities of traditional physical customization are performed by programming the configurable Clos network.
Figures 4A - 4D illustrate a programming embodiment for the configurable Clos network of Figure 2. The example programming shown in Figures 4A - 4D is based on a configurable Clos network having four, 4-port, spine cards and four, 4-port, leaf cards. Four cards with four ports each are used for ease of illustration. Figures 4A - 4D illustrate a top-down, simplified schematic view of a connection segment of the configurable Clos network. In each of Figures 4A - 4D, the spine cards and leaf cards are numbered 0 - 3. The spine cards and leaf cards are connected by a substrate having formed thereon a programmable switch ASIC. Figure 10 illustrates an example of a programmable switch ASIC that may be programmed according to Figures 4A - 4D.
Figure 4A illustrates the programmed connections at the 0th layer of a segment of a configurable Clos network. The leaf cards are designated 0 - 3, the spine cards 0 - 3, and the interconnect substrate and switch ASIC is designated as 0. However, these designations are purely arbitrary. For simplicity of illustration, some details of the connection of the interconnect substrate and switch ASIC are omitted; however, Figure 4A is sufficient to illustrate programming of the 0th switch ASIC.
As can be seen in Figure 4A, port 0 of the 0th leaf card is bi-directionally coupled to port 0 of the 0th spine card, with the dashed lines within the boundaries of the interconnect substrate and switch ASIC representing the signal paths between the leaf cards and the spine cards. Similarly, the Oth ports of the leaf cards 1 , 2, and 3 are coupled, respectively, to the Oth ports of the spine cards 1 , 2, and 3.
Figure 4B illustrates an example of programming of the 1 st switch ASIC. As can be seen, the 1 st port of the Oth leaf card is coupled to the 1 st port of the 1 st spine card, the 1 st port of the 1 st leaf card is coupled to the 1 st port of the 2nd spine card, the 1 st port of the 2nd leaf card is coupled to the 1 st port of the 3rd spine card, and the 1 st port of the 3rd leaf card is coupled to the 1 st port of the Oth spine card.
Figure 4C illustrates an example of programming of the 2nd switch ASIC. As can be seen, the programmed signal paths between the 2nd ports of the leaf and spine cards are shifted by one compared to the programmed signal paths of Figure 4B.
Figure 4D illustrates an example of programming of the 3rd, and last, switch ASIC. As can be seen, the programmed signal paths between the 3rd ports of the leaf and spine cards are shifted by one compared to the programmed signal paths of Figure 4C.
Thus, Figures 4A - 4D illustrate an example of switch ASIC programming in which each port of the leaf cards is coupled at least once to each spine card, and each port of the spine cards is coupled at least once to each leaf card. This signal path shifting may be accomplished using the example mechanisms shown in Figure 10. The signal path shifting may provide a somewhat uniform distribution of signals between leaf and spine cards.
The example programming shown in Figures 4A - 4D is summarized in
Table I.
Table I
INTERCONNECT LEAF-PORT SPINE-PORT
0 0-0 0-0 1 -0 1 -0
2-0 2-0
3-0 3-0
0-1 1 -1
1 -1 2-1
2-1 3-1
3-1 0-1
0-2 2-2
1 -2 3-2
2-2 0-2
3-2 1 -2
3 0-3 3-3
1 -3 0-3
2-3 1 -3
3-3 2-3
Figures 5 - 9 illustrate programming embodiments for the configurable Clos network of Figure 2. Each spine card has 12 ports, and each leaf card has 12 ports. In the absence of the herein disclosed programming and corresponding programmable switches, only when 12 spine cards and 12 leaf cards are installed are all available ports coupled, and all available hardware ports are used.
In Figure 5, the spine cards 1 1 0s are represented by the columns, and the substrates 130/switches 140 are represented by the rows. Thus, in a 12 X 12 switch fabric, there would be twelve switches 140, which are represented by rows 0 - 1 1 , and labeled as substrate. Thus, as can be seen in Figure 5, each crossover between a column and a row represents a connection, and each such connection is given a number, from 0 to 1 1 . Thus, row 0, column 0 has a number 0 registered, meaning the first port of the first connector 150 (and by extension the first leaf card 154) connects to the first port of the first spine card 1 10. In terms of the previously described numbering scheme, the 0th port of the 0th connector 150 (and by extension, the 0th leaf card) connects to the Oth port of the Oth spine card (and vice versa, since the connection paths are bidirectional). Thus, the programming illustrated for the Oth switch 140 provides for straight-through programming. Following row 0 (i.e., the first row) row 1 has the port shifted by one, such that spine card 0 (the first spine card) is the output connection to connector 1 1 (and leaf card 1 1 (the twelfth leaf card)). The process of shifting connections by one with each succeeding row proceeds until all the optical switch positions are programmed for all rows. Thus, the numbers 0 - 1 1 represent the programming of the Clos switch. The effect of the programming shown in Figure 5 is that each port of each spine card connects to each leaf card at least once, and no spine card port is connected more than once to a specific leaf card. Conversely, each port of each leaf card connects to each spine card at least once, and no leaf card port is connected more than once to a specific spine card. Moreover, the programming provides for uniformly distributed signaling with all hardware fully utilized and no unutilized spine or leaf ports.
Figure 6 illustrates a programming scheme when only six spine cards and six leaf cards are installed. In Figure 6, the programmable switch ensures that all available ports are in fact used, even though the number (6 each) of spines and leafs is less than the maximum number (1 2) of spines and leafs. As can be seen in Figure 6, spine card 0 connects to leaf card 0 twice (the 0 at position A and the 0 at position B). In fact, each spine card has exactly two connections to each of the leaf cards. Thus, with the programming of Figure 6, all the hardware ports of each spine card and each leaf card are used. In addition, because there are two connections between each spine card port and each leaf card port, the available bandwidth over the two connections is higher than if only one connection is made.
Figure 7 illustrates an alternate programming scheme that may be used when the configurable Clos network is configured with eight spine cards and eight leaf cards. The 12 inputs or uplinks available on each leaf card now are distributed over 8 spine cards. As can be seen in Figure 7, row (switch 140(0)) 0 is programmed to 0 - 7, row 1 (switch 140(1 )) is programmed to 7 - 6, and so on. The result is that each leaf card has four single and four double connections to the spine cards. The left most, or 0th, leaf card is connected twice to the 0th, 7th, 6th, and 5th spine cards, and once to the 4th, 3rd, 2nd, and 1 st spine cards That is, spine card 0, for example, is connected twice input ports 0, 7, 6, and 5, and once at ports 4, 3, 2, and 1 . Although the programming illustrated in Figure 7 results in a non-symmetric switch fabric, all of the available ports of the leafs and spines are connected and used.
Figure 8 illustrates an alternate programming scheme that may be implemented when the configurable Clos network is configured with 1 2 leaves and 1 0 spines. In Figure 8, row 0 is programmed 0 - 9, row 1 is programmed 9, 0, 1 , ... 8, row 2 is programmed 8, 9, 0, 1 , 2 ... 7, and so on. The result is that each leaf has 8 single and 2 double connections to the spines.
Figure 9 illustrates the programming scheme that may be implemented when four leaf cards and four spine cards are installed, but without the four spine cards being installed in fill order. That is, the four leaf cards occupy positions 0, 2, 3, and 5, and positions 1 , 4, and 6 - 1 1 are empty of leaf cards. In Figure 9, leaf row 0 is programmed 0, 2, 3, and 5; leaf row 1 is programmed 5, 0, 2, and 3, and so on. Thus, in the arrangement of Figure 9, each leaf has 4 triple connections to the spines. The example of Figure 9 shows connectors (leaf cards) occupying the same location numbers as the spines, although such an arrangement between connectors and spines is not necessary.
Figure 10 illustrates an embodiment of a programmable circuit switch used with the configurable Clos network of Figure 2. In Figure 10, small-scale programmable optical switch 140 is implemented as an ASIC on substrate 130. The switch 140 uses moveable, two-sided mirrors 142 that are positioned to route signals between leaf outputs and spine inputs, and spine outputs and leaf inputs, by changing the signal path 90 degrees. The mirrors 142 are angled 45 degrees relative to a path between the outputs and the inputs, and are moveable along this inclined position. The mirrors are positioned in one position using a positive electrostatic force, and in the other position by a negative electrostatic force. For ease of illustration, Figure 1 0 shows the connections for two leaf cards and two spine cards leaf cards (L1 and L2 and spine cards S2 and S3). Mirrors 142(2) and 142(3) are positioned to route the signals from the respective cards. Thus a signal arriving at L10 (leaf out) is diverted by mirror 142(2) to S2| (spine in) and a signal arriving at S20 (spine out) is diverted by mirror 142(2) to L1 | (leaf in), for example.
Using the arrangement of moveable mirrors 142 shown in Figure 10, any output from a leaf (that is, from any LN0) can be routed or directed to any spine input (that is, to any SN|). The embodiment shown in Figure 10 is one example of a programmable circuit switch, and other circuit switch may be used to route signals between the leafs and spines. For example, a switch similar to that shown in Figure 10 can use fixed, two-sided mirrors that are "turned on" to divert a signal path from straight line to 90 degrees, rather than using moveable, two-sided mirrors. In another alternative, instead of a single substrate 1 30 with its switch ASIC 140, two substrates, each with one-way switch ASICs, could be used.
Figure 1 1 is a flow chart illustrating an example of a method for configuring a Clos network. In Figure 1 1 , method 200 begins at block 205 when switch controller 1 60 determines a number of leafs and a number of spines connected to the switch fabric. In block 210, the controller 160 determines a number of available leaf and spine ports. In block 21 5, the controller 160 generates control signals to position the mirrors 142 to provide form a maximum number of leaf-spine optical connections. Method 200 then ends.

Claims

We claim:
1 . A reconfigurable Clos network, comprising: a plurality of spines, each of the spines having a plurality of spine ports; a plurality of leafs, each of the leafs having a plurality of leaf ports; and a plurality of interconnects connecting the spines and the leafs, each of the interconnects comprising: a programmable circuit switch, the circuit switch between the spine ports and the leaf ports, the circuit switch comprising: a plurality of spine inputs; a plurality of spine outputs; a plurality of leaf inputs; a plurality of leaf outputs, wherein the circuit switch operates to transport data between the spine ports and the leaf ports.
2. The configurable Clos network of claim 1 , further comprising a plurality of leaf connectors, each of the leaf connectors coupled to a corresponding one of the plurality of leafs, wherein each leaf is coupled to each of the interconnects, and wherein each of the interconnects is coupled to each of the spines.
3. The reconfigurable Clos network of claim 1 , wherein the circuit switch comprises an electrical switch.
4. The reconfigurable Clos network of claim 1 , wherein the circuit switch comprises an optical switch, and wherein the optical switch comprises two-sided moveable mirrors.
5. The reconfigurable Clos network of claim 4, wherein in a first position, a movable two-sided mirror directs a first signal from a leaf output to a spine input and a second signal from a spine output to a leaf input.
6. The reconfigurable Clos network of claim 5, wherein: a number of leaf ports is less that a number of leafs and a number of spines equals the number of leafs.
7. The reconfigurable Clos network of claim 6, wherein the circuit switch couples each of the ports of each of the leafs to a port of each of the spines and couples each of the ports of each of the spines to a port of each of the leafs.
8. The reconfigurable Clos network of claim 5, wherein: a number of leaf ports equals a number of leafs and a number of spines equals the number of leafs.
9. The reconfigurable Clos network of claim 7, wherein the circuit switch couples each of the ports of each of the leafs to a port of each of the spines and couples each of the ports of each of the spines to a port of each of the leafs.
10. A reconfigurable Clos network, comprising: leafs; spines; a switch fabric that connects the leafs and the spines, wherein the switch fabric couples each leaf port of each leaf to at least one spine port of each spine.
1 1 . The reconfigurable Clos network of claim 10, wherein the switch fabric comprises a plurality of interconnect substrates, and wherein a number of substrates equals a number of leafs.
12. The configurable Clos network of claim 1 1 , wherein each of the configurable switches comprises a mechanism that opens or closes a bidirectional signal path between a leaf and a spine.
13. The configurable Clos network of claim 12, wherein the configurable switches are programmed to couple each of the ports of each of the leafs to a port of each of the spines and to couple each of the ports of each of the spines to a port of each of the leafs.
14. A method for configuring a Clos network, comprising: determining a number of leafs and a number of spines, each of the leafs and each of the spines having a number of ports; and operating a switch to couple each of the ports of each of the leafs to at least one port of each of the spines and to couple each of the ports of each of the spines to at least one port of each of the leafs.
15. The method of claim 14, wherein the number of leafs is less than the number of ports in a leaf, and the number of spines equals the number of leafs.
PCT/US2012/035918 2012-05-01 2012-05-01 Configurable clos network Ceased WO2013165390A1 (en)

Priority Applications (4)

Application Number Priority Date Filing Date Title
CN201280074442.XA CN104380679A (en) 2012-05-01 2012-05-01 Configurable clos network
PCT/US2012/035918 WO2013165390A1 (en) 2012-05-01 2012-05-01 Configurable clos network
US14/398,579 US9584373B2 (en) 2012-05-01 2012-05-01 Configurable Clos network
EP12876066.7A EP2845358A4 (en) 2012-05-01 2012-05-01 CONFIGURABLE CLOSED NETWORK

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/US2012/035918 WO2013165390A1 (en) 2012-05-01 2012-05-01 Configurable clos network

Publications (1)

Publication Number Publication Date
WO2013165390A1 true WO2013165390A1 (en) 2013-11-07

Family

ID=49514656

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/US2012/035918 Ceased WO2013165390A1 (en) 2012-05-01 2012-05-01 Configurable clos network

Country Status (4)

Country Link
US (1) US9584373B2 (en)
EP (1) EP2845358A4 (en)
CN (1) CN104380679A (en)
WO (1) WO2013165390A1 (en)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2016149797A1 (en) * 2015-03-23 2016-09-29 Aeponyx Inc. Photonic switches, photonic switching fabrics and methods for data centers
US9634959B2 (en) 2014-04-08 2017-04-25 International Business Machines Corporation High-density, fail-in-place switches for computer and data networks
GB2549156A (en) * 2015-11-05 2017-10-11 Rockley Photonics Ltd Optoelectronic switch
US9900672B2 (en) 2015-04-24 2018-02-20 Rockley Photonics Limited Optoelectronic switch architectures
US10034069B2 (en) 2015-09-29 2018-07-24 Rockley Photonics Limited Optoelectronic switch
US10491973B2 (en) 2015-04-24 2019-11-26 Rockley Photonics Limited Optoelectronic switch

Families Citing this family (18)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9860116B1 (en) 2013-06-18 2018-01-02 Cisco Technology, Inc. Physical network orchestration for data centers
US9531645B2 (en) 2014-07-29 2016-12-27 Mellanox Technologies Ltd. Cable backplane
CN115082282B (en) 2015-06-10 2025-10-31 无比视视觉技术有限公司 Image processor and method for processing images
US11178072B2 (en) * 2015-06-10 2021-11-16 Mobileye Vision Technologies Ltd. Image processor and methods for processing an image
US9866474B2 (en) * 2015-12-01 2018-01-09 Quanta Computer Inc. Centralized server switch management
CN107959642B (en) * 2016-10-17 2020-08-07 华为技术有限公司 Method, apparatus and system for measuring network paths
US10365445B2 (en) 2017-04-24 2019-07-30 Mellanox Technologies, Ltd. Optical modules integrated into an IC package of a network switch having electrical connections extend on different planes
US10116074B1 (en) * 2017-04-30 2018-10-30 Mellanox Technologies, Ltd. Graded midplane
US20210297757A1 (en) * 2018-07-12 2021-09-23 Panduit Corp. Spatial spectral mesh
US11223574B2 (en) * 2019-12-27 2022-01-11 Google Llc Multi-stage switching topology
US12481795B2 (en) 2023-03-16 2025-11-25 Oracle International Corporation Techniques for validating cloud regions built at a prefab factory
US12568015B2 (en) 2023-03-16 2026-03-03 Oracle International Corporation Techniques for building cloud regions at a prefab factory
US12493457B2 (en) 2023-03-16 2025-12-09 Oracle International Corporation Mobile prefab factory for building cloud regions
US12301465B2 (en) * 2023-05-05 2025-05-13 International Business Machines Corporation Allocating network elements to slices of nodes in a network
US12483530B2 (en) 2023-06-28 2025-11-25 Oracle International Corporation Techniques for rotating network addresses in prefab regions
US12489717B2 (en) 2023-08-04 2025-12-02 Cisco Technology, Inc. Identifying switching paths for ports of crosspoint switch
US12425300B2 (en) 2023-11-27 2025-09-23 Oracle International Corporation Techniques for rotating resource identifiers in prefab regions
US12541355B2 (en) 2024-01-19 2026-02-03 Oracle International Corporation Techniques for image-based region build

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20100097934A1 (en) * 2008-10-21 2010-04-22 Broadcom Corporation Network switch fabric dispersion
US20100254652A1 (en) * 2009-04-01 2010-10-07 Kirkpatrick Peter E Optical Network for Cluster Computing
US20100254703A1 (en) * 2009-04-01 2010-10-07 Kirkpatrick Peter E Optical Network for Cluster Computing
US8060682B1 (en) * 2008-11-14 2011-11-15 Qlogic, Corporation Method and system for multi-level switch configuration

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5875314A (en) * 1996-11-01 1999-02-23 Northern Telecom Limited Configurable connection fabric for providing serial backplanes with adaptive port/module bandwidth
US6658177B1 (en) * 1999-12-13 2003-12-02 Memlink Ltd. Switching device and method of fabricating the same
JP2002214546A (en) * 2000-11-15 2002-07-31 Oki Electric Ind Co Ltd Optical switch
WO2006026648A2 (en) 2004-08-31 2006-03-09 Herman Miller, Inc. Reconfiguring control relationships among devices
US7389046B1 (en) 2005-01-18 2008-06-17 Woven Systems, Inc. Simplified switching interconnect using wavelength division multiplexing
US7391952B1 (en) 2006-08-31 2008-06-24 Corning Cable Systems Llc Pre-connectorized fiber optic cable network interconnection apparatus
US7983194B1 (en) 2008-11-14 2011-07-19 Qlogic, Corporation Method and system for multi level switch configuration
US7756371B1 (en) 2009-01-30 2010-07-13 Corning Cable Systems Llc Optical fiber interconnection devices and systems using same

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20100097934A1 (en) * 2008-10-21 2010-04-22 Broadcom Corporation Network switch fabric dispersion
US8060682B1 (en) * 2008-11-14 2011-11-15 Qlogic, Corporation Method and system for multi-level switch configuration
US20100254652A1 (en) * 2009-04-01 2010-10-07 Kirkpatrick Peter E Optical Network for Cluster Computing
US20100254703A1 (en) * 2009-04-01 2010-10-07 Kirkpatrick Peter E Optical Network for Cluster Computing

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP2845358A4 *

Cited By (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10389654B2 (en) 2014-04-08 2019-08-20 International Business Machines Corporation High-density, fail-in-place switches for computer and data networks
US9634959B2 (en) 2014-04-08 2017-04-25 International Business Machines Corporation High-density, fail-in-place switches for computer and data networks
US10749817B2 (en) 2014-04-08 2020-08-18 International Business Machines Corporation High-density, fail-in-place switches for computer and data networks
US11039228B2 (en) 2015-03-23 2021-06-15 Aeponyx Inc. Photonic switches, photonic switching fabrics and methods for data centers
WO2016149797A1 (en) * 2015-03-23 2016-09-29 Aeponyx Inc. Photonic switches, photonic switching fabrics and methods for data centers
US11310569B2 (en) 2015-03-23 2022-04-19 Aeponyx Inc. Photonic switches, photonic switching fabrics and methods for data centers
CN107710702A (en) * 2015-03-23 2018-02-16 艾易珀尼斯公司 Photonic switching system, photonic switching fabric and the method for data center
US11206466B2 (en) 2015-03-23 2021-12-21 Aeponyx Inc. Photonic switches, photonic switching fabrics and methods for data centers
US10694268B2 (en) 2015-03-23 2020-06-23 Aeponyx Inc. Photonic switches, photonic switching fabrics and methods for data centers
US11190859B2 (en) 2015-03-23 2021-11-30 Aeponyx Inc. Photonic switches, photonic switching fabrics and methods for data centers
US10757489B2 (en) 2015-03-23 2020-08-25 Aeponyx Inc. Photonic switches, photonic switching fabrics and methods for data centers
US10880627B2 (en) 2015-03-23 2020-12-29 Aeponyx Inc. Photonic switches, photonic switching fabrics and methods for data centers
US10491973B2 (en) 2015-04-24 2019-11-26 Rockley Photonics Limited Optoelectronic switch
US9900672B2 (en) 2015-04-24 2018-02-20 Rockley Photonics Limited Optoelectronic switch architectures
US10034069B2 (en) 2015-09-29 2018-07-24 Rockley Photonics Limited Optoelectronic switch
GB2549156B (en) * 2015-11-05 2021-09-15 Rockley Photonics Ltd Optoelectronic switch
GB2549156A (en) * 2015-11-05 2017-10-11 Rockley Photonics Ltd Optoelectronic switch

Also Published As

Publication number Publication date
EP2845358A1 (en) 2015-03-11
US9584373B2 (en) 2017-02-28
EP2845358A4 (en) 2015-12-16
US20150110488A1 (en) 2015-04-23
CN104380679A (en) 2015-02-25

Similar Documents

Publication Publication Date Title
US9584373B2 (en) Configurable Clos network
AU2011305638B2 (en) Transpose box based network scaling
AU2011305575B2 (en) Transpose boxes for network interconnection
CA2831559C (en) Incremental high radix network scaling
US9762481B2 (en) Efficient highly connected deployment units
EP1087637B1 (en) Compact high-capacity switch
CN103608797A (en) Network transpose box and switch operation based on backplane Ethernet
CN102740177A (en) Non-blocking expandable multistage photoswitch array and working method thereof
Parsons et al. High radix all-optical switches for software-defined datacentre networks
US8160061B2 (en) Redundant network shared switch
US7473851B2 (en) Cross-connecting by permutations using configurable printed circuits
Kabacinski et al. Rearranging algorithms for log 2 (N, 0, p) switching networks with even number of stages
EP0739146A2 (en) Telecommunications switch
US7912019B1 (en) Applications of upgradeable scalable switching networks
US20240171886A1 (en) Fabric modules for high-radix networks
JPH04273739A (en) Unit switch module and atm cross-connect device using the same

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 12876066

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

WWE Wipo information: entry into national phase

Ref document number: 14398579

Country of ref document: US

REEP Request for entry into the european phase

Ref document number: 2012876066

Country of ref document: EP

WWE Wipo information: entry into national phase

Ref document number: 2012876066

Country of ref document: EP