CN114363241B - High-dimensional Torus network architecture and self-adaptive routing method - Google Patents

High-dimensional Torus network architecture and self-adaptive routing method Download PDF

Info

Publication number
CN114363241B
CN114363241B CN202210015846.7A CN202210015846A CN114363241B CN 114363241 B CN114363241 B CN 114363241B CN 202210015846 A CN202210015846 A CN 202210015846A CN 114363241 B CN114363241 B CN 114363241B
Authority
CN
China
Prior art keywords
node
dimension
data packet
routing
packet
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.)
Active
Application number
CN202210015846.7A
Other languages
Chinese (zh)
Other versions
CN114363241A (en
Inventor
向东
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.)
Tsinghua University
Original Assignee
Tsinghua University
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 Tsinghua University filed Critical Tsinghua University
Priority to CN202210015846.7A priority Critical patent/CN114363241B/en
Publication of CN114363241A publication Critical patent/CN114363241A/en
Application granted granted Critical
Publication of CN114363241B publication Critical patent/CN114363241B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L45/00—Routing or path finding of packets in data switching networks
    • H04L45/12—Shortest path evaluation
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L45/00—Routing or path finding of packets in data switching networks
    • H04L45/14—Routing performance; Theoretical aspects
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L45/00—Routing or path finding of packets in data switching networks
    • H04L45/18—Loop-free operations
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L49/00—Packet switching elements
    • H04L49/90—Buffering arrangements
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D30/00—Reducing energy consumption in communication networks
    • Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks

Landscapes

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

Abstract

The invention provides a high-dimensional Torus network architecture and a self-adaptive routing method, wherein the high-dimensional Torus network architecture comprises the following components: multiple unit dimensions arranged from a low dimension to a high dimension in sequence; the dimension of the multi-cell includes k n A computing unit, k n The node coordinates of each calculation unit are respectively k 0 ，k 1 ，……，k n‑1 ；k n The calculation units are according to 0,1,. And k n -1 are connected in sequence to form a closed loop; k is a radical of n And the multiple-unit dimension is a positive integer greater than or equal to 2, when the routing is performed and the current node coordinate of the transmitted data packet is not equal to the target node coordinate, the multiple-unit dimension corresponds to i =0 to n-1: if d is i ‑c i >k i A/2 or-k i /2≤d i ‑c i <0, selecting the next calculation unit as a routing node in the negative direction of the ith dimension; if 0<d i ‑c i <k i A ratio of/2 to d i ‑c i <‑k i And/2, selecting the next computing unit as a routing node along the positive direction of the ith dimension, so that the efficiency of the routing method in the Torus network is improved, and the equal-division bandwidth is improved.

Description

High-dimensional Torus network architecture and self-adaptive routing method
The application is a divisional application of a patent application named as 'high-dimensional Torus network architecture and self-adaptive routing method', the application date of the original application is 2019, 05 and 15, and the application number is 201910408017.3.
Technical Field
The invention relates to the technical field of distributed computing, in particular to a high-dimensional Torus network architecture and a self-adaptive routing method.
Background
Interconnection Networks (interconnects) are a bus technology that breaks through the performance limits that are reached due to electrical limitations, and are a common solution to the system level communication problem of modern digital systems. An interconnection network is an important module which affects the performance and the expansibility of the whole multiprocessor system, and at present, how to provide low-delay and high-efficiency communication between processing nodes of the interconnection network based on the prior art becomes a key problem to be solved urgently by the interconnection network.
The interconnection network is described primarily by three elements: topology, switching mechanism and routing method. High performance interconnects are the core technology for building supercomputers. High performance interconnects have evolved mainly from two directions: (1) A low-degree network, wherein the degree of each computing node (or processor) is low, and the low-dimensional Torus network is mainly adopted; gemini interconnects such as Cray Titan with a node degree of 8; IBM Blue/Gene L and P, node degree 6; (2) High-Radix network (High-Radix), mainly comprises fat tree, dragonfly network, etc. Wherein, the low degree network is characterized in that: the throughput rate of each physical channel is high, but the switch design of each node is simple and the network diameter is high; higher-order networks are characterized by low network diameters, but complex switch designs and relatively low throughput per physical channel.
The Torus network is an interconnection network, which is a completely symmetrical topology structure and has many excellent characteristics, such as the same node degree, simple structure, multiple paths, good expandability and the like, and thus, the Torus network is widely applied to commercial systems. The routing method performance in the Torus network still needs to be improved; and the bisection bandwidth is low.
Disclosure of Invention
The invention aims to provide a high-dimensional Torus network architecture and a self-adaptive routing method, which can improve the system performance and solve the problems of low efficiency and low equal-divided bandwidth of the routing method in the Torus network in the prior art.
In order to achieve the purpose, the invention provides the following technical scheme:
a high-dimensional Torus network architecture, comprising:
multiple unit dimensions arranged from a low dimension to a high dimension in sequence; wherein the multi-unit dimension includes k n A computing unit, k n The node coordinates of each calculation unit are respectively k 0 ，k 1 ，……，k n-1 ；k n The calculation units are according to 0,1,. And k i -1 are connected in sequence to form a closed loop; k is a radical of n Is a positive integer greater than or equal to 2, i represents the number of dimensions;
the dimension of the multi-unit is inCurrent node coordinates c of transmitted data packets when routing is performed i Coordinates d with target node i When not equal, corresponding to i =0 to n-1:
if d is i -c i >k i A/2 or-k i /2≤d i -c i <0, selecting the next calculation unit as a routing node in the negative direction of the ith dimension;
if 0<d i -c i <k i /2 or d i -c i <-k i 2, selecting the next calculation unit as a routing node along the positive direction of the ith dimension;
when the next computing unit selects the routing node, the preset safety requirement set by the flow control technology is met; the preset safety requirements are as follows: if the next computing unit includes f idle virtual channels and s secure virtual channels, then: if f >1, or f =1 and s ≧ 1, or f =1 and s =0 and the p × cube method is followed from the current computation unit to the next computation unit, the next computation unit serves as a routing node.
Optionally, the dimension of the high-dimensional Torus network architecture is greater than or equal to 2.
In order to achieve the purpose, the invention also provides the following technical scheme:
an adaptive routing method is applied to the high-dimensional Torus network architecture, and comprises the following steps:
step one, in an n-dimension Torus network architecture, when a data packet route is set, the coordinate of a current node is C (C) 0 ，c 1 ，...，c i ，…，c n-1 ) Destination node coordinate is D (D) 0 ，d 1 ，…，d i ，…，d n-1 ) The number f of idle cache buffers is (f) 0 +，f 0 -，f 1 +，f 1 -，...，f i +，f i -，…，f n-1 +，f n-1 -) the number of secure data packets S is (S) 0 +，s 0 -，s 1 +，s 1 -，...，s i +，s i -，…，s n-1 +，s n-1 -)；
Wherein, c i Representing the coordinates of the node in which the message is currently located in the i-th dimension, d i Representing the coordinates of the destination node of the message in dimension i, f i +、f i -representing the number of free virtual channels in the input ports of the nodes adjacent to the current node C in the i-dimension positive and negative directions, respectively, s i +、s i -representing the number of secure virtual channels in the input ports of the neighboring nodes of the current node C in the positive and negative directions, respectively, of the ith dimension;
step two, if the coordinate C of the node where the data packet is located is equal to the coordinate D of the destination node, the data packet reaches the destination node, and the routing is finished;
step three, if the coordinate C of the node where the data packet is located at present is not equal to the coordinate D of the destination node, sequentially executing the following operations from the dimension i =0 to the dimension i = n-1:
if the number of nodes calculated for dimension i is 3, if d i -c i <0 and | d i -c i I =1, CH = CH & { CH = CH | > i - }; CH is an optional output channel set of the data packet, and the initial value is a null set phi; if d is i -c i <0 and | d i -c i I =2, CH = CH & { CH = g & |) i + }; if d is i -c i >0 and | d i -c i I =1, CH = CH & { CH = CH | > i + }; if d is i -c i >0 and | d i -c i I =2, CH = CH & { CH = g & |) i -}；
If the number of nodes in dimension i is 4, if d i -c i <0 and | d i -c i | ≦ 2, then CH = CH ≦ U { CH i - }; if d is i -c i <0 and | d i -c i |>2, then CH = CH { CH } { (CH } CH } U i + }; if d is i -c i >0 and | d i -c i | ≦ 2, then CH = CH ≦ U { CH i + }; if d is i -c i >0 and | d i -c i |>2, then CH = CH & { CH + i -}；
If the calculation node of dimension i is larger than 4, if d i -c i >k i A/2 or-k i /2≤d i -c i <0, and f i -，s i -providing a jump ch of packets along the current point C to the negative direction of its i dimension if the conditions of the flow control technique driven by the preset dynamic packet state are met i -go, CH = CH ═ c { CH } i -}；
If 0<d i -c i ≤k i /2 or d i -c i <-k i 2, and f i +，s i + providing the jump ch from the current point C to the ith dimension positive direction if the preset dynamic packet state driven flow control technique condition is satisfied i + forward, CH = CH { [ CH ] } { (CH {) } { (CH } { (CH) } and i +}；
and step four, adaptively selecting one channel from the set CH as a final output channel CH.
Optionally, the adaptive routing method further includes: the shortest path to the destination node is provided for the packet.
Optionally, the providing a shortest path to the destination node for the data packet specifically includes:
step S1: defining a safety data packet and a non-safety data packet, and if the data packet meets one of the following conditions, determining that the data packet is the safety data packet:
if the next jump of the data packet is through the surrounding channel, the number of the calculation nodes placed in the dimensionality is larger than 4, and the dimensionality of the surrounding channel is the lowest dimensionality of the data packet which needs to pass through the surrounding channel from the current point to the target node, and for the node on the next jump, the data packet is a safety data packet;
if the data packet reaches the destination node from the current node, the number of the calculation nodes placed in any dimension is no longer required to pass through a surrounding channel with the number being more than 4, and the next jump of the data packet follows a p-cube method, the data packet is a safety data packet for the nodes on the next jump;
otherwise, the data packet is an unsafe data packet;
step S2: under the virtual cut-through VCT switching technology, an input port of any node router comprises two buffer areas, each buffer area at least completely stores one data packet, and the two buffer areas are not classified;
setting f to represent the number of idle virtual channels in the input port, and s to represent the number of safe virtual channel resources in the input port; the state of the buffer zone comprises idle, safe and non-safe, and the idle virtual channel represents that no data packet is stored in the buffer zone; the safe virtual channel represents that a buffer zone stores a safe data packet or the buffer zone resource originally stores an unsafe data packet, and the unsafe data packet applies for the buffer zone resource on the next hop node; the non-secure virtual channel represents that a buffer area stores a non-secure data packet, and the non-secure data packet does not apply for buffer area resources on a next hop node;
and step S3: based on the buffer state information provided in step S2, it is set that the next-hop router node has f idle virtual channels and S secure virtual channels, and the following flow control mechanism is provided:
if f >1, the packet may proceed to the next node;
if f =1 and s ≧ 1, the packet can proceed to the next node;
if f =1 and s =0, the packet may advance to the next node if the p × cube method is followed from the current point to the next node;
otherwise, the packet cannot proceed to the next node.
Optionally at the target node d i And current node c i When the difference is only one hop, adopting a p-cube route, which specifically comprises the following steps:
will be the current node (c) 0 ，c 1 ，...，c i ，…，c n-1 ) And a target node (d) 0 ，d 1 ，…，d i ，…，d n-1 ) Division of the differently numbered dimensions into two subsets U 1 And U 2 Initial U 1 And U 2 Setting as an empty set phi; for all dimensions i =0 to i = n-1, if c i ≠d i And the next jump of the dimension is ch i +, then U 1 ←U 1 U { i }; otherwise, the next hop for that dimension is ch i -，U 2 ←U 2 ∪{i}；
From whenCoordinates C (C) of the front node 0 ，c 1 ，…，c n-1 ) To destination node coordinate D (D) 0 ，d 1 ，…，d n-1 ) Data packet of, spanning U 2 All dimensions in (1) when U 2 After all dimensions are walked, U is adaptively selected 1 Transmit the data packet.
According to the specific embodiment provided by the invention, the invention discloses the following technical effects: the invention provides a high-dimensional Torus network architecture and a self-adaptive routing method n And each computing unit adopts an adaptive routing method when carrying out routing selection, selects an optimal path based on a shortest path principle, considers the number of the computing units of the current network and selects the shortest path by changing the transmission direction. The designed high-dimensional torus network can effectively improve the equal bandwidth and the data transmission efficiency of the network and is beneficial to improving the network performance.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art that other drawings can be obtained according to these drawings without creative efforts.
FIG. 1 is a schematic diagram of a high dimensional Torus network architecture according to the present invention;
FIG. 2 is a schematic diagram of a 3 × 3 × 3 × 3 × 3 5 dimensional torus network according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of a 5-dimensional hypercube subnetwork extracted from a 3 × 3 × 3 × 3 network provided by an embodiment of the present invention;
FIG. 4 is a flow chart of the adaptive routing method of the present invention;
fig. 5 is a schematic diagram of a switch implementing dynamic packet status information based according to an embodiment of the present invention;
FIG. 6-1 is a schematic illustration of an equally divided bandwidth along a given dimension of a high dimensional torus network with an even number of compute nodes, according to an exemplary embodiment of the present invention;
FIG. 6-2 is a schematic diagram of an equally divided bandwidth along a given dimension of a high dimensional torus network with an odd number of compute nodes according to an example embodiment of the present invention;
FIGS. 7-1 and 7-2 are schematic diagrams of the calculation of the divided bandwidth along the two highest dimensions according to the embodiment of the present invention;
fig. 7-3 are diagrams of an equal-divided bandwidth calculation along the d0-d7 dimensions provided by an example of an embodiment of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The invention aims to provide a high-dimensional Torus network architecture and a self-adaptive routing method, which effectively improve the equal division bandwidth and the data transmission efficiency of a network and are beneficial to improving the network performance.
In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in further detail below.
Fig. 1 and fig. 2 are schematic diagrams illustrating a Torus network architecture according to an embodiment of the present invention.
Each dx in fig. 1 represents a network architecture of one dimension.
As shown in fig. 2, the high-dimensional Torus network architecture of the present invention includes: multiple unit dimensions arranged from a low dimension to a high dimension in sequence; wherein the multi-unit dimension includes k n A computing unit, k n The node coordinates of each calculation unit are respectively k 0 ，k 1 ，……，k n-1 ；k n The calculation units are according to 0,1,. And k i -1 are connected in sequence to form a closed loop; k is a radical of n Is a positive integer greater than or equal to 2, and i represents the number of dimensions. Specifically, the dimension of the high-dimensional Torus network architecture is greater than or equal to 2. In the embodiment of the present invention, the calculation unit 0 → the calculation unit 1 → the calculation unit k n -1 are connected in sequence.
As a specific implementation, the high dimensional Torus network architecture comprises: n three-unit dimensions which are sequentially arranged from a low dimension to a high dimension, wherein N is a positive integer greater than or equal to 7; the three-unit dimension comprises three calculation units, and the node coordinates of the three calculation units are 0,1,2 respectively; the three computing units are sequentially connected according to the sequence of 0 → 1 → 2 to form a closed loop.
When the high-dimensional Torus network architecture transmits data packets, routing selection is needed, and equally dividing the bandwidth is an important parameter for measuring the quality of a network. If the number of the computing nodes arranged in different dimensions is different, the halved bandwidths of the same network along different dimensions are different; specifically, if the transmission is performed along the dimension where more computing nodes are placed, the equally divided bandwidth is smaller; the transmission is equally divided along the dimension in which fewer compute nodes are placed, and the equal bandwidth is larger.
The routing method determines the transmission path of each message or grouped data packet in the network and sends the data packet to the destination node correctly without errors. Typically, messages or packets always provide the same path between any pair of nodes, regardless of network state, and the method is simple, but the path is typically not the optimal path. For a three-unit dimension, in order to improve the equally divided bandwidth, a shortest path should be selected for data packet transmission, specifically, when the three-unit dimension performs routing, and the current node coordinate c of the transmitted data packet i Coordinates d with target node i Inequality, i.e., when a packet is not transmitted to the destination node, there are several situations:
if d is i -c i = -1, then transmit the data packet along the negative direction of the dimension, i.e. select the next calculation unit as the routing node in the direction opposite to the first direction.
Wherein, as shown in figure 2Shown as d i -c i Scenario of =1 includes two combinations, combination (1) d i =0 and c i =1, or combination (2) d i =1 and c i And (2). For combination (1) d i =0 and c i If =1, the transmission direction of the packet is 1 → 0, and the shortest path is to be transmitted in the opposite direction to the first direction, the next calculation unit is selected as the routing node. For combination (2) d i =1 and c i And =2, in this case, the transmission direction of the data packet is 2 → 1, and the shortest path is to be transmitted in the opposite direction of the first direction, and the next calculation unit is selected as the routing node.
d i -c i Scenario of =2 includes a combination, namely combination (3) d i =2 and c i =0, in this case the transmission direction of the data packet is 0 → 2, and the shortest path should be the transmission in the opposite direction to the first direction, and the next calculation unit is selected as the routing node.
If d is i -c i =1 or d i -c i And =2, selecting the next calculation unit as the routing node in the first direction. In particular, d i -c i Scenario of =1 includes two combinations (4) d i =1 and c i =0, or combination (5) d i =2 and c i =1; combination (4) d i =1 and c i If =0, the transmission direction of the data packet is 0 → 1, the shortest path should be to the first direction, the next calculation unit is selected as the routing node, and the combination (5) and the combination (4) are the same, which is not described herein again.
d i -c i Scenario of =2 includes combination (6) d i =0 and c i =2, when the transmission direction of the data packet is 2 → 0, the shortest path should be to transmit to the first direction, and the next calculation unit is selected as the routing node.
As a specific implementation mode, the high-dimensional Torus network architecture comprises at least three-unit dimensions, the three-unit dimensions comprise three computing units, each computing unit adopts an adaptive routing method when routing selection is carried out, an optimal path is selected based on a shortest path principle, the number of the computing units of the current network is considered, the shortest path is selected by changing the transmission direction, and the designed high-dimensional Torus network can effectively improve the equal bandwidth and the data transmission efficiency of the network, and is favorable for improving the network performance. The embodiment of the invention solves the problems of low efficiency and low equal division bandwidth of a routing method in a Torus network in the prior art.
Optionally, the high dimensional Torus network architecture further comprises: n three-unit dimensions are sequentially arranged from a low dimension to a high dimension, wherein N is a positive integer greater than or equal to 7. When the high-dimensional Torus network architecture is 10-dimensional, N is 7, at this time, 3 computing units are configured for d0-d9, and the diameter of the network architecture is 10; the total number of computing nodes of the system reaches 3 10 =59049. When the high-dimensional Torus network architecture is 11-dimensional, N is 8, and at the moment, 3 computing units are configured for d0-d10, wherein the diameter of the network architecture is 11; the total number of computing nodes of the system reaches 3 11 =177147. When the high-dimensional Torus network architecture is 12-dimensional, N is 8, and at the moment, 3 computing units are configured for d0-d11, and the diameter of the network architecture is 12; the total number of computing nodes of the system reaches 3 12 ＝531441。
As an embodiment of the present invention, the high-dimensional Torus network architecture further includes: the three-unit dimension, the three k-unit dimension and the four-unit dimension are sequentially arranged from the low dimension to the high dimension; the k unit dimensions comprise k computing units, k is a positive integer greater than or equal to 5, namely d4 is also a three-unit dimension, d5, d6 and d7 are k unit dimensions, d8 is a four-unit dimension, and the high-dimensional Torus network architecture is an 8-dimensional network; when k is equal to 5, the diameter of the high-dimensional Torus network architecture is only 12, and the total number of computing nodes of the system adopting the configuration is (3) 3 )*3*(5 3 ) 4 + 4=40500, exceeding the computing power of class E computers.
As an embodiment of the present invention, the high-dimensional Torus network architecture further includes: four three-unit dimensions and two k-unit dimensions which are sequentially arranged from a low dimension to a high dimension; wherein the k unit dimension comprises k computing units, k is a positive integer greater than or equal to 5, namely d0-d7 are all configured with three computing units, and d8D9, configuring k unit dimensions, wherein the high-dimension Torus network architecture is an 8-dimension network; when k is equal to 5, the diameter of the high-dimensional Torus network architecture is only 12, and the total number of computing nodes of the system adopting the configuration is (3) 7 )*(5 2 )*4＝164025。
Optionally, when performing routing, the four-unit dimension performs the following adaptive routing method: if the current node coordinate c of the transmitted data packet i With target node coordinates d i Is not equal, then
(1) If d is i -c i <0 and | d i -c i Less than or equal to 2 or d i -c i And =3, selecting the next calculation unit as the routing node in the opposite direction of the first direction. Wherein d is i -c i <0 and | d i -c i In the case of | ≦ 2, i.e. in the first direction, c i At d i In the above, possible combinations include: d is a radical of i =0 and c i ＝1；d i =0 and c i ＝2；d i =1 and c i ＝2；d i =1 and c i ＝3；d i =2 and c i ＝3；d i =2 and c i ＝4；d i =3 and c i =4; in the above cases c i Return to original path d i And selecting the next calculation unit as a routing node in the opposite direction of the first direction. d i -c i Case of =3, including d i =3 and c i =0, the next calculation unit should be selected as the routing node in the opposite direction of the first direction.
(2) If d is i -c i And =1 or 2, selecting the next calculation unit as the routing node in the first direction. In d i -c i Case =1 or 2, i.e. in the first direction, d i At c i In front of (2); d i -c i Combinations that may occur when =1 include: d is a radical of i =1 and c i ＝0；d i =2 and c i ＝1；d i =3 and c i ＝2；d i -c i Combinations that may occur when =2 include: d i =2 and c i ＝0；d i =3 and c i =1; in the above cases, the next calculation unit is selected as the routing node to the first direction.
In the above embodiment of the present invention, when performing routing selection, the following adaptive routing method is executed in the multi-unit dimension: current node coordinates c of transmitted data packets i Coordinates d with target node i When not equal, corresponding to i =0 to n-1:
(1) If d is i -c i >k i A/2 or-k i /2≤d i -c i <0, then the next calculation unit is selected as the routing node in the negative direction of the ith dimension. Wherein d is i -c i >k i At/2, i.e. in the positive direction, d i At c i And the number of nodes of the distance between the two in the positive direction is greater than k i A/2, the number of nodes in the reverse direction is less than k i 2, therefore, the shortest path is used for transmitting the data packet to the reverse direction; -k i /2≤d i -c i <Similarly, the embodiment of the present invention is not described herein again.
(2) If 0<d i -c i <k i /2 or d i -c i <-k i And/2, selecting the next calculation unit as a routing node along the positive direction of the ith dimension. Wherein, d i -c i <k i At/2, i.e. in the positive direction, d i At c i And the number of nodes of the distance between the two in the positive direction is less than k i A/2, the number of nodes in the reverse direction is greater than k i 2, so the packet is transmitted in the forward direction as the shortest path; d i -c i <-k i Similarly, the embodiment of the present invention is not described herein again.
And when the next computing unit selects the routing node, the buffer requirement set by the flow control technology is met.
Specifically, the preset safety requirements are:
if the next computing unit includes f idle virtual channels and s secure virtual channels, then:
if f >1, or f =1 and s ≧ 1, or f =1 and s =0 and the p × cube method is followed from the current compute unit to the next compute unit, the next compute unit can serve as a routing node.
Regarding the preset security requirement, firstly, a security data packet and a non-security data packet are defined, and if the data packet meets one of the following conditions, the data packet is the security data packet:
1) If the next jump of the data packet is to pass through a surrounding channel in the unit dimension, the number of calculation units placed in the dimension is more than 4, the dimension of the surrounding channel is the lowest dimension of the data packet which needs to pass through the surrounding channel from the current node to the destination node, and the data packet is a safety data packet for the node on the next jump.
2) If the data packet does not need to pass through a surrounding channel with the number of the calculation nodes being more than 4 dimensions from the current node to the destination node any more, and the next jump of the data packet follows a deadlock-free adaptive routing method in a high-dimensional torus network, such as a p-cube routing method, the data packet is a safe data packet for the node on the next jump.
If the data packet does not satisfy both of the above conditions, the data packet is an unsecured data packet.
In addition, regarding the Virtual channels, under a Virtual-Cut Through (VCT) switching technology, the input port of any node router includes two Virtual channels, each Virtual channel includes a buffer, each buffer can completely store a data packet, and the two buffers are not classified. Let f denote the number of free virtual channels in the input port, and s denote the number of secure virtual channel resources in the input port. The buffer area can be divided into three states of idle, safe and non-safe, and the idle virtual channel indicates that no data packet is stored in the buffer area. And if the number of the idle virtual channels is more than 1, the next calculation unit can be used as a routing node.
Specifically, the secure virtual channel indicates that the buffer stores a secure data packet, or the buffer resource originally stores an insecure data packet, but the insecure data packet has already applied for the buffer resource on the next hop node; the non-secure virtual channel indicates that the buffer stores the non-secure data packet, and the non-secure data packet has not been applied to the buffer resource on the next-hop node. And if the number of the idle virtual channels is equal to 1 and the number s of the safe virtual channels is more than or equal to 1, the next computing unit can be used as a routing node.
Each virtual channel has a corresponding Last bit (LT Token, LT), when the header flit of a packet reaches a virtual channel, the LT bit of the virtual channel is set to 1, and the LT bit of another virtual channel is set to 0. The LT is used to ensure in-order transmission of data packets, and if two virtual channels in the input port store header flits with the same destination node, the previous header flit (the virtual channel corresponding to LT 0) enters the virtual channel allocation stage. Otherwise, both data packets enter the virtual channel allocation stage. In addition, each virtual channel of each input port has a SAFE bit for indicating whether a data packet occupying the virtual channel is a secure data packet, and the bit is updated when the virtual channel receives a data packet header flit.
When the header flit is transmitted to a downstream node, the SAFE bit will transmit update information to the upstream node. When the upstream node receives the update information, the router of the upstream node updates the number SAFE of the secure virtual channels counted in the counter, and the FREE parameter is also updated correspondingly. When the non-secure data packet stored in the non-secure virtual channel applies for the virtual channel of the downstream routing node, the non-secure virtual channel is dynamically changed into a secure virtual channel, and corresponding SAFE update information is transmitted to the upstream routing node for the upstream node to update the SAFE domain.
Further, if f =1 and s =0, and a p × cube method is followed from the current calculation unit to the next calculation unit, the next calculation unit may serve as a routing node.
Regarding the p-cube routing method, the p-cube routing method is a partially adaptive row routing method for an ultra-cubic network, and for a target point (d) 0 ，d 1 ，...，d n-1 ) And current point (c) 0 ，c 1 ，...，c n-1 ) When all differ by only one hop, the p-cube routing method can be used.
Dividing the dimension with different labels between the current node and the target node into two subsets U 1 And U 2 Initial U 1 And U 2 Setting as an empty set phi; for all dimensions i =0 to i = n-1, if c i ≠d i And the next jump of the dimension is ch i +, then U 1 ←U 1 U { i }; otherwise, the next hop for that dimension is ch i -，U 2 ←U 2 ∪{d i }。
The p-cube routing method is as follows: from (c) 0 ，c 1 ，...，c n-1 ) To (d) 0 ，d 1 ，...，d n-1 ) First adaptively span U 2 All dimensions in (1) when U 2 After all dimensions are walked, selecting U adaptively 1 Transmit the data packet.
The high-order torus network fully adaptive routing algorithm is constructed by using a p-cube routing algorithm as a reference routing strategy and a flow control technology based on dynamic data packet state information.
By way of example, referring to fig. 3, fig. 3 is a schematic diagram of the interior of a router fabric, i.e., implementing switches based on dynamic packet state information. Compared with the traditional router structure, a FREE domain and an SAFE domain are required to be added at the output port of the router; the FREE field indicates the number of idle virtual channels of the input port of the downstream node, and the SAFE field indicates the number of the SAFE data packets stored in the buffer zone of the input port of the downstream node. When the data packet advances, the two parameters are utilized, and the condition that two virtual channels of any input port simultaneously store the non-secure data packet can be effectively avoided according to the preset security requirement, namely if f is larger than 1, the data packet can advance to the next node; if f =1 and s ≧ 1, the packet can proceed to the next node; if f =1 and s =0, the packet may proceed to the next node if it is a secure packet to the next node; otherwise, the packet cannot proceed to the next node.
Optionally, in the above embodiment of the present invention, after selecting the next computing unit as the routing node, the high-dimensional Torus network architecture outputs an optional channel set;
the selectable channel set comprises selectable channels of the current computing unit and selectable channels of the next computing unit.
Wherein if transmission is made in a first (forward) direction, the next calculation unit is selected as routing node, ch i ＝CH i ∪{ch i +}，ch i An alternative channel set for the current computing unit, c i + is the set of selectable channels, ch, for the next computational unit in the first direction i Is an optional set of channels.
If the transmission is in the reverse direction of the first direction, the next calculation unit is selected as a routing node, CH i ＝CH i ∪{c i -}，CH i An optional set of channels for the current computational unit, c i -an optional set of channels, CH, being the next calculation unit in the opposite direction of the first direction i Is an optional set of channels.
The embodiment of the invention relates to an interconnection architecture of a high-dimensional torus network, which is suitable for a VCT or cache wormhole (BWH) exchange technology, and the diameter of the network is only 12 when the number of computing nodes reaches 40500 through network layout; when the number of the computing nodes reaches 164025, the network diameter is still only 12, and the total computing capacity of the system is about 4.5 times of that of an E-class computer at the moment; when the number of nodes reaches 177147, the network diameter is only 11, and the computing power of a 4.8E-level computer is achieved; when the number of the nodes reaches 531441, the network diameter is only 12, and the computing capacity of an E-class computer is 14.4 times.
With the foregoing description of the high-dimensional Torus network architecture provided by the embodiment of the present invention, an adaptive routing method provided by the embodiment of the present invention will be described below with reference to the accompanying drawings.
As shown in fig. 4, the adaptive routing method of the present invention is applied to the high-dimensional Torus network architecture, and the adaptive routing method includes:
step 401, when setting up data packet routing in the n-dimension Torus network architecture,the coordinate of the current node is C (C) 0 ，c 1 ，...，c i ，…，c n-1 ) Destination node coordinate is D (D) 0 ，d 1 ，…，d i ，…，d n-1 ) The number f of idle cache buffers is (f) 0 +，f 0 -，f 1 +，f 1 -，...，f i +，f i -，…，f n-1 +，f n-1 -) the number of secure data packets S is (S) 0 +，s 0 -，s 1 +，s 1 -，...，s i +，s i -，…，s n-1 +，s n-1 -)；。
Wherein, c i Representing the coordinates of the node in which the message is currently located in the i-th dimension, d i Representing the coordinates of the destination node of the message in the i-dimension, f i +、f i -representing the number of free virtual channels in the input ports of the nodes adjacent to the current node C in the i-dimension positive and negative directions, respectively, s i +、s i -representing the number of secure virtual channels in the input ports of the neighboring nodes of the current node C in the positive and negative directions, respectively, of the ith dimension.
Step 402, if the coordinate C of the node where the data packet is currently located is equal to the coordinate D of the destination node, that is, C = D, the data packet has already reached the destination node, and the routing is ended.
Step 403, if the current node coordinate C of the data packet is not equal to the destination node coordinate D, sequentially performing the following operations from dimension i =0 to dimension i = n-1:
if the number of nodes calculated for dimension i is 3, if d i -c i <0 and | d i -c i I =1, CH = CH & { CH = CH | > i - }; CH is an optional output channel set of the data packet, and the initial value is an empty set phi; if d is i -c i <0 and | d i -c i I =2, CH = CH ═ CH { [ CH ] } then i + }; if d is i -c i >0 and | d i -c i | =1, then CH = CH { [ CH ] } CH { (1) } and i + }; if d is i -c i >0 and | d i -c i I =2, CH = CH & { CH = g & |) i -}；
If the calculation section of dimension iNumber of points 4, if d i -c i <0 and | d i -c i | ≦ 2, then CH = CH ≦ U { CH i - }; if d is i -c i <0 and | d i -c i |>2, then CH = CH & { CH + i + }; if d is i -c i >0 and | d i -c i | ≦ 2, then CH = CH ≦ U { CH i + }; if d is i -c i >0 and | d i -c i |>2, then CH = CH & { CH + i -}；
If the calculation node of dimension i is larger than 4, if d i -c i >k i A/2 or-k i /2≤d i -c i <0, and f i -，s i -providing a jump ch of packets along the current point C to the negative direction of its i dimension if the conditions of the flow control technique driven by the preset dynamic packet state are met i -go, CH = CH ═ c { CH } i -}；
If 0<d i -c i ≤k i /2 or d i -c i <-k i 2, and f i +，s i + providing the jump ch from the current point C to the ith dimension positive direction if the preset dynamic packet state driven flow control technique condition is satisfied i + forward, CH = CH { [ CH ] } { (CH {) } { (CH } { (CH) } and i +}。
step 404, obtaining a set CH of output channels selectable by the data packet through the operation of step 403, and adaptively selecting one channel from the set CH as a final output channel CH.
Further, the adaptive routing method further includes: the shortest path to the destination node is provided for the packet. The method specifically comprises the following steps:
step S1: defining a safety data packet and a non-safety data packet, and if the data packet meets one of the following conditions, determining that the data packet is the safety data packet:
if the next jump of the data packet is through the surrounding channel, the number of the calculation nodes placed in the dimension is larger than 4, the dimension where the surrounding channel is located is the lowest dimension of the data packet which needs to pass through the surrounding channel from the current point to the destination node, and the data packet is a safety data packet for the node on the next jump.
If the data packet reaches the destination node from the current node, and the number of the calculation nodes placed in any dimension is no longer needed to pass through the surrounding channel with the number of the calculation nodes larger than 4, and the next jump of the data packet follows a deadlock-free self-adaptive routing method (such as p-cube routing) in the high-dimension torus network, the data packet is a safe data packet for the node on the next jump.
If the data packet does not meet the two conditions, the data packet is an insecure data packet;
step S2: under the virtual cut-through VCT switching technology, an input port of any node router comprises two buffer areas, each buffer area at least completely stores one data packet, and the two buffer areas are not classified.
Setting f to represent the number of idle virtual channels in the input port, and s to represent the number of secure virtual channel resources in the input port; the buffer area can be divided into three states of idle, safe and non-safe; the idle virtual channel represents that no data packet is stored in the buffer area; the safe virtual channel represents that a buffer area stores a safe data packet or the buffer area resource originally stores an unsafe data packet, and the unsafe data packet applies for the buffer area resource on the next hop node; the non-secure virtual channel indicates that the buffer stores the non-secure data packet, and the non-secure data packet has not been applied to the buffer resource on the next-hop node.
And step S3: based on the buffer state information provided in step S2, it is set that the next-hop router node has f idle virtual channels and S secure virtual channels, and the following flow control mechanism is provided:
if f >1, the packet may proceed to the next node.
If f =1 and s ≧ 1, the packet can proceed to the next node.
If f =1 and s =0, the packet may advance to the next node if the p-cube method is followed from the current point to the next node.
Otherwise, the packet cannot proceed to the next node.
The secure virtual channel may originally store a secure data packet, or may be converted from a non-secure virtual channel state, which may be implemented only when the non-secure data packet applies for a next-hop buffer resource, and this is also the core of the flow control technology based on the dynamic data packet state.
Optionally, a target node d i And current node c i When the difference is only one hop, adopting a p-cube route, which specifically comprises the following steps:
will be the current node (c) 0 ，c 1 ，...，c i ，…，c n-1 ) And a target node (d) 0 ，d 1 ，…，d i ，…，d n-1 ) Division of the differently numbered dimensions into two subsets U 1 And U 2 Initial U 1 And U 2 Setting as an empty set phi; for all dimensions i =0 to i = n-1, if c i ≠d i And the next jump of this dimension is ch i +, then U 1 ←U 1 ∪{d i }; otherwise, the next hop of the dimension is ch i -，U 2 ←U 2 ∪{d i }；
From the coordinate C (C) of the current node 0 ，c 1 ，…，c n-1 ) To destination node coordinate D (D) 0 ，d 1 ，…，d n-1 ) Data packet of, spanning U 2 All dimensions in (1), when U 2 After all dimensions are walked, U is adaptively selected 1 Transmit the data packet.
In the embodiment of the invention, when the high-dimensional Torus network architecture is used for routing, each computing unit adopts an adaptive routing method when the high-dimensional Torus network architecture is used for routing, an optimal path is selected based on a shortest path principle, the number of the computing units of the current network is considered, the shortest path is selected by changing the transmission direction, and the designed high-dimensional Torus network can effectively improve the equal bandwidth and the data transmission efficiency of the network and is beneficial to improving the network performance. The embodiment of the invention solves the problems of low efficiency and low equal division bandwidth of the routing method in the Torus network in the prior art.
By way of example, FIG. 5 shows a five-dimensional hypercube sub-cube extracted from a 3-ary 5-cube. Reliable communication may be directed based on the security information extracted from the sub-cube.
The divided bandwidth of a network is an important criterion for the quality of the network.
FIGS. 6-1 through 6-2 illustrate the calculation of the halved bandwidth when divided along a given dimension of the high dimensional torus network.
FIG. 6-1 shows the case when the dimension contains an even number of compute nodes. For one k 0 ×k 1 ×...×k i-1 ×k i ×k i+1 ×...×k n-1 The network, the divided bandwidth along the ith dimension can be calculated as follows: 4k 0 k 1 ...k i-1 k i+1 ...k n-1 。
FIG. 6-2 shows the case where the ith dimension contains an odd number of compute nodes. In this case, it is necessary to compute the equal division of the bandwidth at multiple levels, assuming k 0 ≤k 1 ≤k 2 ≤...≤k n-2 ≤k n-1 Then the halved bandwidth of the n-dimensional torus network can be calculated as follows:
BB＝2k 0 k 1 ...k i-1 k i+1 ...k n-2 +2k 0 k 1 ...k i-1 k i+1 ...k n-3 +...+2k 0 k 1 +2k 0 +2
fig. 7-1 through 7-3 illustrate calculating the halved bandwidth of a 3 × 3 × 3 × 3 × 3 × 3 × 5 × 5 network, respectively. Fig. 7-1, 7-2 show the bisection bandwidth calculations along the top two dimensions, respectively.
Fig. 7-3 are the bisection bandwidth calculations along the d0-d7 dimension.
Fig. 3 shows details of a corresponding switch implementation of the present invention.
The quality of a network is judged to have a plurality of characteristic parameters: node degree, diameter, and equally divided bandwidth. As shown in table 1, the 3 × 3 × 3 × 3 × 5 × 5 × 5 × 4 network has a higher bandwidth equal to that of the 16 × 16 × 16 × 12 × 2 network k-Computer,2 × 3 × 2 × 16 × 16 × 32, and IBM Blue/Gene P,64 × 48 × 32 adopted by IBM pool/Gene Q, and the network diameter is greatly reduced.
Table 1:
Figure BDA0003460692970000171
the equal division bandwidth of the high-dimensional torus network 3 × 3 × 3 × 3 × 3 × 3 × 3 × 5 × 5 is better than that of the dragonfly network (28, 392) (i.e., each router group contains 28 routers and totally contains 392 router groups), and the degree of the routers in each dragonfly network is 56, while the degree of the nodes in the 3 × 3 × 3 × 3 × 3 × 3 × 3 × 5 × 5 network is 20 degrees, so that the bandwidth of each physical channel is approximately 3 times that of the (28, 392) dragonfly network.
The above-described embodiments of the apparatus are merely illustrative, and the units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of the present embodiment. One of ordinary skill in the art can understand and implement it without inventive effort.
Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented by software plus a necessary general hardware platform, and certainly can also be implemented by hardware. With this understanding in mind, the above-described technical solutions may be embodied in the form of a software product, which can be stored in a computer-readable storage medium such as ROM/RAM, magnetic disk, optical disk, etc., and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the methods described in the embodiments or some parts of the embodiments.
The principles and embodiments of the present invention have been described herein using specific examples, which are provided only to help understand the method and the core concept of the present invention; meanwhile, for a person skilled in the art, according to the idea of the present invention, the specific embodiments and the application range may be changed. In view of the foregoing, the description is not to be taken in a limiting sense.

Claims (5)

1. A high-dimensional Torus network architecture, characterized in that the high-dimensional Torus network architecture comprises:
multiple unit dimensions arranged from a low dimension to a high dimension in sequence; wherein the multi-unit dimension includes k n A computing unit, k n The node coordinates of each calculation unit are respectively k 0 ，k 1 ，……，k n-1 ；k n The calculation units are according to 0,1,. And k i -1 are connected in sequence to form a closed loop; k is a radical of n Is a positive integer greater than or equal to 2, i represents the number of dimensions;
the current node coordinate c of the transmitted data packet when the multi-unit dimension performs routing i Coordinates d with target node i When not equal, corresponding to i =0 to n-1:
if d is i -c i >k i A/2 or-k i /2≤d i -c i <0, selecting the next calculation unit as a routing node in the negative direction of the ith dimension;
if 0<d i -c i <k i /2 or d i -c i <-k i 2, selecting the next calculation unit as a routing node along the positive direction of the ith dimension;
when the next computing unit selects the routing node, the preset safety requirement set by the flow control technology is met; the preset safety requirements are as follows: if the next computing unit includes f idle virtual channels and s secure virtual channels, then: if f >1, or f =1 and s ≧ 1, or f =1 and s =0 and a p × cube method is followed from the current computing unit to the next computing unit, the next computing unit serves as a routing node;
the p-cube routing method is a partial adaptive row routing method for the ultra-cubic network, and is used for a target node (d) 0 ，d 1 ，...，d n-1 ) And the current node (c) 0 ，c 1 ，...，c n-1 ) All differ by only one hop, adopt p* -cube routing method;
dividing the dimension with different labels between the current node and the target node into two subsets U 1 And U 2 Initial U 1 And U 2 Setting as an empty set phi; i =0 to i = n-1 for all dimensions, if c i ≠d i And the next jump of this dimension is ch i +, then U 1 ←U 1 U { i }; otherwise, the next hop for that dimension is ch i -，U 2 ←U 2 ∪{d i }；
The p-cube routing method is as follows: from the coordinate C (C) of the current node 0 ，c 1 ，...，c n-1 ) To destination node coordinate D (D) 0 ，d 1 ，...，d n-1 ) First adaptively span U 2 All dimensions in (1) when U 2 After all dimensions are walked, selecting U adaptively 1 Transmit the data packet.
2. The high-dimensional Torus network architecture of claim 1, wherein the dimension of said high-dimensional Torus network architecture is greater than or equal to 2.
3. An adaptive routing method applied to the high-dimensional Torus network architecture according to any one of claims 1 to 2, the adaptive routing method comprising:
step one, in an n-dimension Torus network architecture, when a data packet route is set, the coordinate of a current node is C (C) 0 ，c 1 ，...，c i ，…，c n-1 ) Destination node coordinate is D (D) 0 ，d 1 ，…，d i ，…，d n-1 ) The number f of idle cache buffers is (f) 0 +，f 0 -，f 1 +，f 1 -，...，f i +，f i -，…，f n-1 +，f n-1 -) the number of secure data packets S is (S) 0 +，s 0 -，s 1 +，s 1 -，...，s i +，s i -，…，s n-1 +，s n-1 -)；
Wherein, c i Representing the coordinates of the node in which the message is currently located in the i-th dimension, d i Representing the coordinates of the destination node of the message in the i-dimension, f i +、f i -representing the number of free virtual channels in the input ports of the nodes adjacent to the current node C in the i-dimension positive and negative directions, respectively, s i +、s i -representing the number of secure virtual channels in the input ports of the neighboring nodes of the current node C in the i-th dimension positive and negative directions, respectively;
step two, if the coordinate C of the node where the data packet is located is equal to the coordinate D of the destination node, the data packet reaches the destination node, and the routing is finished;
step three, if the coordinate C of the node where the data packet is located at present is not equal to the coordinate D of the destination node, sequentially executing the following operations from the dimension i =0 to the dimension i = n-1:
if the number of nodes calculated for dimension i is 3, if d i -c i <0 and | d i -c i I =1, CH = CH & { CH = CH | > i - }; CH is an optional output channel set of the data packet, and the initial value is a null set phi; if d is i -c i <0 and | d i -c i I =2, CH = CH & { CH = g & |) i + }; if d is i -c i >0 and | d i -c i I =1, CH = CH & { CH = CH | > i + }; if d is i -c i >0 and | d i -c i I =2, CH = CH & { CH = g & |) i -}；
If the number of nodes calculated for dimension i is 4, if d i -c i <0 and | d i -c i | ≦ 2, then CH = CH ≦ U { CH i - }; if d is i -c i <0 and | d i -c i |>2, then CH = CH { CH } { (CH } CH } U i + }; if d is i -c i >0 and | d i -c i | ≦ 2, then CH = CH ≦ U { CH i + }; if d is i -c i >0 and | d i -c i |>2, then CH = CH { CH } { (CH } CH } U i -}；
If the calculation node of dimension i is larger than 4, if d i -c i >k i A/2 or-k i /2≤d i -c i <0, and f i -，s i -providing a jump ch of the packet along the negative direction from the current point C to its i-dimension if the conditions of the flow control technique driven by the predefined dynamic packet states are met i -go, CH = CH ═ c { CH } i -}；
If 0<d i -c i ≤k i /2 or d i -c i <-k i 2, and f i +，s i + providing the jump ch from the current point C to the ith dimension positive direction if the preset dynamic packet state driven flow control technique condition is satisfied i + forward, CH = CH { [ CH ] } { (CH {) } { (CH } { (CH) } and i +}；
step four, a channel is selected from the set CH in a self-adaptive mode to serve as a final output channel CH;
at the target node (d) 0 ，d 1 ，...，d n-1 ) And the current node (c) 0 ，c 1 ，...，c n-1 ) When all the routes are different by one hop, the method adopts a p-cube route, and specifically comprises the following steps:
dividing the dimension with different labels between the current node and the target node into two subsets U 1 And U 2 Initial U 1 And U 2 Setting as an empty set phi; for all dimensions i =0 to i = n-1, if c i ≠d i And the next jump of the dimension is ch i +, then U 1 ←U 1 U { i }; otherwise, the next hop for that dimension is ch i -，U 2 ←U 2 ∪{d i }；
The p-cube routing method is as follows: from the coordinate C (C) of the current node 0 ，c 1 ，...，c n-1 ) To destination node coordinate D (D) 0 ，d 1 ，...，d n-1 ) First adaptively span U 2 All dimensions in (1), when U 2 After all dimensions are finished, the adaptive selection U is carried out 1 Transmit the data packet.
4. The adaptive routing method according to claim 3, further comprising:
the shortest path to the destination node is provided for the packet.
5. The adaptive routing method according to claim 4, wherein the providing the shortest path to the destination node for the packet specifically comprises:
step S1: defining a safety data packet and a non-safety data packet, and if the data packet meets one of the following conditions, determining that the data packet is the safety data packet:
if the next jump of the data packet passes through the surrounding channel, the number of the calculation nodes placed in the dimension is larger than 4, the dimension where the surrounding channel is located is the lowest dimension of the data packet which needs to pass through the surrounding channel from the current point to the destination node, and the data packet is a safety data packet for the node on the next jump;
if the data packet reaches the destination node from the current node, the number of the calculation nodes placed in any dimension is no longer required to pass through a surrounding channel with the number being more than 4, and the next jump of the data packet follows a p-cube method, the data packet is a safety data packet for the nodes on the next jump;
otherwise, the data packet is an insecure data packet;
step S2: under the virtual cut-through VCT switching technology, an input port of any node router comprises two buffer areas, each buffer area at least completely stores one data packet, and the two buffer areas are not classified;
setting f to represent the number of idle virtual channels in the input port, and s to represent the number of safe virtual channel resources in the input port; the state of the buffer zone comprises idle, safe and non-safe, and the idle virtual channel represents that no data packet is stored in the buffer zone; the safe virtual channel represents that a buffer area stores a safe data packet or the buffer area resource originally stores an unsafe data packet, and the unsafe data packet applies for the buffer area resource on the next hop node; the non-secure virtual channel represents that a buffer zone stores a non-secure data packet, and the non-secure data packet does not apply for buffer zone resources on a next hop node;
and step S3: based on the buffer state information provided in step S2, it is set that the next-hop router node has f idle virtual channels and S secure virtual channels, and the following flow control mechanism is provided:
if f is greater than 1, the data packet can advance to the next node;
if f =1 and s ≧ 1, the packet can proceed to the next node;
if f =1 and s =0, the packet may advance to the next node if the p-cube method is followed from the current point to the next node;
otherwise, the packet cannot proceed to the next node.
CN202210015846.7A 2019-05-15 2019-05-15 High-dimensional Torus network architecture and self-adaptive routing method Active CN114363241B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210015846.7A CN114363241B (en) 2019-05-15 2019-05-15 High-dimensional Torus network architecture and self-adaptive routing method

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910408017.3A CN110198268A (en) 2019-05-15 2019-05-15 The high-dimensional Torus network architecture and adaptive routing method
CN202210015846.7A CN114363241B (en) 2019-05-15 2019-05-15 High-dimensional Torus network architecture and self-adaptive routing method

Related Parent Applications (1)

Application Number Title Priority Date Filing Date
CN201910408017.3A Division CN110198268A (en) 2019-05-15 2019-05-15 The high-dimensional Torus network architecture and adaptive routing method

Publications (2)

Publication Number Publication Date
CN114363241A CN114363241A (en) 2022-04-15
CN114363241B true CN114363241B (en) 2023-02-28

Family

ID=67752788

Family Applications (2)

Application Number Title Priority Date Filing Date
CN202210015846.7A Active CN114363241B (en) 2019-05-15 2019-05-15 High-dimensional Torus network architecture and self-adaptive routing method
CN201910408017.3A Pending CN110198268A (en) 2019-05-15 2019-05-15 The high-dimensional Torus network architecture and adaptive routing method

Family Applications After (1)

Application Number Title Priority Date Filing Date
CN201910408017.3A Pending CN110198268A (en) 2019-05-15 2019-05-15 The high-dimensional Torus network architecture and adaptive routing method

Country Status (1)

Country Link
CN (2) CN114363241B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114745345B (en) * 2022-04-07 2023-10-20 中国人民解放军国防科技大学 Efficient routing method and device for dynamic back pressure adjustment of network on chip
CN117319308B (en) * 2023-11-28 2024-02-23 苏州元脑智能科技有限公司 Data transmission method, device, equipment, system and storage medium

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102170402A (en) * 2011-05-31 2011-08-31 清华大学 A deadlock-free adaptive routing algorithm in a Torus network
CN104539536A (en) * 2014-12-01 2015-04-22 清华大学 Flow control of dynamic state driving and Torus network adaptive routing method
CN107959643A (en) * 2017-12-29 2018-04-24 曙光信息产业(北京)有限公司 A kind of exchange system and its routing algorithm built by exchange chip
CN109561034A (en) * 2018-12-25 2019-04-02 中科曙光信息产业成都有限公司 Three-dimensional network topological structure and its routing algorithm

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101267398B (en) * 2008-04-18 2010-09-01 清华大学 Non deal lock self-adapted routing method in 2-D torus network
NO2776466T3 (en) * 2014-02-13 2018-01-20
CN104079491B (en) * 2014-07-07 2018-04-27 中国科学院计算技术研究所 A kind of router and method for routing towards high-dimensional network
CN105224501B (en) * 2015-09-01 2018-10-02 华为技术有限公司 The method and apparatus improved annulus torus network and its determine data packet transmission path
CN107612746B (en) * 2017-10-12 2020-12-22 曙光信息产业股份有限公司 Torus network construction method, Torus network and routing algorithm

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102170402A (en) * 2011-05-31 2011-08-31 清华大学 A deadlock-free adaptive routing algorithm in a Torus network
CN104539536A (en) * 2014-12-01 2015-04-22 清华大学 Flow control of dynamic state driving and Torus network adaptive routing method
CN107959643A (en) * 2017-12-29 2018-04-24 曙光信息产业(北京)有限公司 A kind of exchange system and its routing algorithm built by exchange chip
CN109561034A (en) * 2018-12-25 2019-04-02 中科曙光信息产业成都有限公司 Three-dimensional network topological structure and its routing algorithm

Also Published As

Publication number Publication date
CN114363241A (en) 2022-04-15
CN110198268A (en) 2019-09-03

Similar Documents

Publication Publication Date Title
US11362934B2 (en) Method to route packets in a distributed direct interconnect network
JP4676463B2 (en) Parallel computer system
Bermond et al. Broadcasting and gossiping in de Bruijn networks
Glass et al. The turn model for adaptive routing
JP6093867B2 (en) Non-uniform channel capacity in the interconnect
US8432830B1 (en) Multi-neighbor proportional forwarding in a network
Wang et al. NovaCube: A low latency Torus-based network architecture for data centers
WO2012162988A1 (en) Efficient adaptive deadlock-free routing algorithms for torus networks
CN114363241B (en) High-dimensional Torus network architecture and self-adaptive routing method
Aydogan et al. Adaptive source routing in multistage interconnection networks
CN110650141B (en) SDN (software defined network) segmented route defense method for link flooding attack
CN117135059B (en) Network topology structure, construction method, routing algorithm, equipment and medium
US20240154906A1 (en) Creation of cyclic dragonfly and megafly cable patterns
CN110324249B (en) Dragonfly network architecture and multicast routing method thereof
Mahapatra et al. Limited multi-path routing on extended generalized fat-trees
Park et al. Generic methodologies for deadlock-free routing
Liang et al. Beyond the performance of three-tier fat-tree: equality topology with low diameter
Stunkel Commercially viable MPP networks
CN113271267B (en) Interconnection network, adaptive routing method, apparatus, electronic device, and storage medium
Jaroš Evolutionary Design of Collective Communications on Wormhole Networks
KR101160508B1 (en) Apparatus for choosing mpls routing route of network centric warfare c4i battleground-moving c4i security communication network and method thereof
Chi-Hsiu et al. Performance evaluation of multi-exaflops machines using Equality network topology
Kiasari et al. A comparative performance analysis of n-cubes and star graphs
Shamaei et al. A Simple and Efficient Fault-Tolerant Adaptive Routing Algorithm for Meshes
Boppana et al. A comparison of adaptive packet-routing algorithms for tori

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant