CN103346908B - A kind of communication referee method towards high-performance calculation - Google Patents
A kind of communication referee method towards high-performance calculation Download PDFInfo
- Publication number
- CN103346908B CN103346908B CN201310254247.1A CN201310254247A CN103346908B CN 103346908 B CN103346908 B CN 103346908B CN 201310254247 A CN201310254247 A CN 201310254247A CN 103346908 B CN103346908 B CN 103346908B
- Authority
- CN
- China
- Prior art keywords
- communication
- arbitration
- node
- information stream
- communication resources
- 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
Links
Abstract
A kind of communication referee method towards high-performance calculation that invention provides, comprises the steps: (1) communication resources logical partitioning;(2) node loads and expands arbitration information stream;(3) described arbitration information stream is resolved;(4) respond according to analysis result in described (3) and service.The communication referee method that the present invention provides, the demand that uniformly communication (Uniform pattern) communicates can be met with non-homogeneous (Non uniform pattern) simultaneously, realize the real-time reasonable distribution of communication resources, high-effect utilize communication resources (bandwidth) to realize high performance communication;The referee method proposed is particularly suited for the emerging high-speed radio mutual contact mode such as optic communication, mm wave RF communication, the utilization rate of the communication resource (bandwidth) is greatly improved, reduce communication routing cost, improve communication speed, can reduce overhead while meeting system performance requirements, " 4P " that meet HPC market requires (performance performance, usefulness productivity, power consumption power and price price).
Description
Technical field
The present invention relates to high-performance computing sector, be specifically related to a kind of communication referee method.
Background technology
It is super that high-performance calculation HPC (High Performance Computing) refers to that utilization has
The high-performance computer of strong computing capability resolves current super large, superelevation, hypercomplex calculating task
.HPC has become as the third-largest important after " theoretical " and " test " of scientific technological advance
Scientific research method, and at space science, process engineering, geoscience, bio information, high energy thing
The key areas such as national science and technology, national defence, industry and finance such as reason play the most important work
With.
Dynamical communication interconnection be the key realizing high-performance calculation, design efficient internet and
Communication mechanism has become as the major issue improving high-performance computer performance.High-performance meter
Calculation generally faces large-scale calculations field, generally by hundreds of rack, and thousands of node structures
Become a huge calculating system, and each node be generally integrated multiple, tens even go up
Hundred processor cores.Traditional mutual contact mode can not meet more to come between rack and node
The harshest high-performance calculation requirement.
Along with application demand be continuously increased and technology development make rapid progress, more and more study by
Novel interconnection resources, such as optical interconnection, mm wave RF interconnection etc. is applied to high-performance calculation interconnection
Network, meets day by day harsh the wanting of the performance (time delay, handling capacity) to network and power consumption
Ask.But these emerging mutual contact modes yet suffer from certain challenge in technique and the aspect such as integrated,
The limited bandwidth resources that can be provided by.These bandwidth resources are the most efficiently utilized to realize efficient communication
Interconnection, is that by the key of high-performance calculation.
At present lot of domestic and international scholar on sheet between high-speed radio interconnection, sheet optical interconnection etc. propose one
A little topological structure and implementations etc., but there is no a kind of effective communication referee method, it is possible to knot
Close the common denominator of these emerging interconnections, limited bandwidth resource allocation is distributed real-time and efficiently
To the communication needed most between, it is achieved high-performance communication interconnect.
Summary of the invention:
In view of this, the main object of the present invention is, for " locality " feature of high-performance calculation
And the communicating requirement of " high-performance low-power-consumption ", it is provided that a kind of communication towards high-performance calculation is secondary
Sanction method, this communication referee method can realize the real-time reasonable distribution of communication resources, meet simultaneously
Uniformly communication is logical with non-homogeneous
News demand, high-effect utilizes these emerging communication resources (bandwidth) to realize high performance communication;
Meet HPC market " 4P " require (performance performance, usefulness productivity,
Power consumption power and price price).
In order to achieve the above object, the present invention provides a kind of communication referee method towards high-performance calculation
, it is characterised in that comprise the steps:
(1) communication resources logical partitioning;
(2) node loads and expands arbitration information stream;
(3) described arbitration information stream is resolved;
(4) respond according to analysis result in described (3) and service.
Wherein, described communication resources refer to realize channel or the information transmission bandwidth resource of information transmission;
In (3) described in described basis, analysis result responds and services, and refers to that arbitrating winning node exists
In the execution cycle, the communication resources obtained are enjoyed exclusive communication right, utilizes obtained to lead to
News resource carries out message transmission.
Described referee method, is further characterized in that, in described step (1), and described communication resources logic
Division refers to, described communication resources are logically divided into arbitration communication resources and data communication
Resource, wherein arbitration communication resources are used for transmitting or carry described arbitration information stream, data communication
Resource is for the transmission of data message.
Described referee method, is further characterized in that, in described step (2), node loads arbitration information
Stream and expansion arbitration information stream are two relatively independent behaviors, and allow behavior described in the two
Executed in parallel or non-executed in parallel;Described expansion arbitration information stream refers to that described node adds logical
The solicited message of news resource, referee method as described in claim 1 and 3, it is characterised in that institute
State expansion arbitration information stream and refer to that described node adds the solicited message to described communication resources, institute
State solicited message and carry out record by bit flag position, and at least include 1 bit information position.
Described referee method, is further characterized in that, described node adds the request to described communication resources
During information, also can increase Informationflow Control information, carry out record by increasing bit flag position.
Described referee method, is further characterized in that, after once arbitration terminates, can be expanded by concept transfer
Fill the precedence of arbitration information stream, change priority arbitration.
Advantages of the present invention
A kind of communication referee method towards high-performance calculation that the present invention provides, it is possible to simultaneously meet all
Even communication (Uniform pattern) and non-homogeneous (Non-uniform pattern) lead to
The demand of letter, it is achieved communication
The real-time reasonable distribution of resource, high-effect utilizes communication resources (bandwidth) to realize high performance communication
;The referee method suitability proposed is strong, it is possible to be applied to optic communication, mm wave RF communication
Etc. emerging high-speed radio mutual contact mode, the utilization rate of the communication resource (bandwidth) is greatly improved,
Play high-effect.Reduce communication routing cost, improve communication speed, it is possible in satisfied systematicness
Can reduce overhead while requiring, " 4P " that meet HPC market requires (performance perfo
Rmance, usefulness productivity, power consumption power and price price).
Accompanying drawing illustrates:
Fig. 1 referee method flow chart according to embodiments of the present invention;
Fig. 2 is communication channel schematic diagram based on RF-I;
Fig. 3 is the basic applicating example of the referee method according to the embodiment of the present invention;
Fig. 4 a referee method expanded application citing one according to embodiments of the present invention;
Fig. 4 b referee method expanded application citing two according to embodiments of the present invention;
Fig. 5 is to describe expanded application according to the arbitration stream information in the referee method of the embodiment of the present invention to lift
Example.
Detailed description of the invention:
For making the object, technical solutions and advantages of the present invention clearer, referring to the drawings,
In conjunction with an emerging communication mutual contact mode--radio frequency interconnection RF-I (Radio Frequency I
Nterconnections), the specific implementation method of the present invention is described further.
Fig. 1 is the referee method flow chart according to the embodiment of the present invention, in order to become apparent from detailed statement originally
A kind of communication referee method towards high-performance calculation that invention proposes, communicates below in conjunction with RF-I
Mode introduces detailed description of the invention and step.Fig. 2 is communication channel schematic diagram based on RF-I, will
Its communication resource (bandwidth) is divided into two logic of class communication channels, arbitrates channel (arbitrat
Ion channel) and data channel (data channels), wherein arbitration channel is used for
Transmitting arbitration information stream, data channel is used for data information.Data channel can be according to reality
Need to be divided into multiple data channel.Communication node obtains the use to data channel by arbitration
Power.
Arbitration information stream had not been empty, from a certain node before any node expands arbitration information
Set out, process node according to communication requirement to arbitration stream information expand communication resource request letter
Breath, is generally represented flag information by bit, represents that this node has data as expanded bit " 1 "
Send, need to make
With data channel, participating in communication resource competition at this arbitration cycle, representing if expanding bit " 0 "
This node no data sends, and is not involved in communication resource competition.
Node, while expanding communication resource request information, loads current arbitration information stream and preserves
In this locality.The node having communication resource request resolves the arbitration information stream loaded, according to arbitration letter
Number of request in breath stream, calculates whether there is data available passage, if also data channel can be used,
Then show that this node is won in this arbitration, it is possible to exclusive 1 data passage enters within the execution cycle
Row data transmission.
The basic applicating example that Fig. 3 is the referee method according to the embodiment of the present invention, for convenience
, hereafter arbitrate with a kind of communication towards high-performance calculation that stream arbitration acute pyogenic infection of finger tip is proposed by the invention
Method.In this example, four nodes are had to compete 2 data channels.Assume this arbitration cycle, node
A, B, D have data to send, and need to use data channel.
Arbitration information stream, from node A, is empty at the beginning, node expanding resource solicited message, because
Its participate in communication resource competition, therefore node A expand bit " 1 ", current arbiter flow of information from
Space-variant is " 1 ... ";Node-B receiver arbitration information stream " 1 .... ", and expand its arbitration information
, because B is also required to compete data channel, therefore its expanding data " 1 ", present node stream
By " 1 .... " be extended for " 11 .... ";Node C need not compete data channel, and therefore it expands
Make up the number according to " 0 ", present node stream by " 11 .... " be extended for " 110 .... ";By that analogy
, node D expand after, arbitration information stream be " 1101 .... ".
Owing to node A is first node expanding arbitration information stream, and data available port number is 2,
Therefore node A wins in this arbitration, it is possible to exclusive 1 data passage communicates;Node B receives
The arbitration information stream arrived for " 1 .... ", show before had 1 node to occupy 1 data
Passage, also remains 1 data available passage, and therefore B also wins in this arbitration, it is possible to exclusive 1
Data channel communicates;The arbitration information stream that node D receives for " 110 .... ", show before
Having had 2 nodes to occupy 2 data passages, without data available passage, therefore D is secondary at this
In sanction, failure, can not monopolize data channel within this arbitration execution cycle and communicate.
Be exemplified as most basic stream arbitration implementation method.Arbitration information can be carried out according to demand
Expansion and perfect, it is achieved communication stream controls and low power loss communication.Fig. 4 is to implement according to the present invention
The expanded application citing of the referee method of example, the communication that in this extension example, perfect node expands is secondary
Sanction information, not only contains communication resource request bit, further comprises flow control bit, purpose
Node ID information.Fig. 5 is in this expanded application example in referee method according to embodiments of the present invention
Arbitration stream information describes, as it is shown in figure 5, node is when expanding arbitration information, will expand 2+log2N number of bit information, wherein the 1st bit is flow control bit, and the reception such as this node has cached
Full, then it is set to " 1 ", showing can not be in this arbitration
The execution cycle receives data message, otherwise sets reset;2nd bit is expressed the communication resource
Request, the same with in basic application examples, as this node participates in resource contention, then set set
, " otherwise 0 ";Last log2N number of bit describes destination node ID, if i.e. this node obtains
Victory, destination node ID of its message transmitted, shared number of bits is log2N, wherein N is for being
Communication section in system is counted.Such as, if this arbitration has 4 communication nodes, then arbitration stream letter
Breath includes 2+log24 bits, i.e. 4 bits, latter two bit describe letter
Breath destination node ID.
In expanded application example, it will be assumed that 4 nodes, contention 1 data passage.Assume current
Arbitration cycle, arbitration information stream needs to utilize data channel to send number from node A, node A
It is believed that breath is required for contention to node C, node B and D and utilizes data channel to send data message to joint
Point C, node C do not have data to send, but its data buffer storage is the fullest, it is impossible to connect from data channel
Receive new data.In this expanded application example, with the addition of communication flow control information, therefore need
Could load, preserve and to resolve complete arbitration letter after all nodes expansion arbitration stream information
Breath, thus the requests for arbitration state obtaining all nodes (is held in the 2nd cycle in this application example
Row loads and resolves, it is ensured that the arbitration stream information resolved is complete).As shown in Figure 4, exist
In first cycle, each node expands its arbitration information, according to arbitration stream information mark as shown in Figure 5
Will position, node A, B, C and D modulation bit stream " 0110 " respectively, " 0100 ", "
1000 " and " 0100 ".At second period, each node resolves complete arbitration information
Stream " 0110010010000100 ", node A is first request data channel request node, enjoys
There is limit priority, therefore have the right to obtain 1 data passage, but the caching of its destination node C is
Full (C: " 1000 "), therefore node A still failure in this is arbitrated;The purpose of node B
Node C caching is effective, and also data available passage, and therefore node B wins in this is arbitrated
.Node D resolves arbitration information stream and knows, node B has been obtained for unique 1 data available and leads to
Road, therefore node D also arbitrates failure.After arbitration terminates, the node B of triumph will monopolize it and obtain
Data channel data information, node A and node D will next time arbitrate in again participate in competing
Strive.
Stream referee method proposed by the invention, can expand priority mechanism, and change node expands secondary
Cut out the priority of stream information, it is ensured that the fairness of arbitration.This application example uses backward round
The method of priority, i.e. A, B, C, D each node backward obtains limit priority, as currently
Arbitration cycle node A priority is the highest, A expand arbitration information stream at first, and B, C take second place, and D is excellent
First level is minimum, the most finally expands arbitration information stream.Next arbitration cycle node D priority is best,
A, B take second place, and the priority of C is minimum, later each arbitration cycle by that analogy, so in certain journey
The fairness of stream referee method proposed by the invention is ensure that on degree.
Above-described embodiment is the preferred Application Example of the present invention, and protection scope of the present invention is not
It is confined to this embodiment.Any amendment that is all in the range of the spirit and principles in the present invention, that made
, improvement, equivalent etc., should be included within the scope of the present invention.
Claims (5)
1. the communication referee method towards high-performance calculation, it is characterised in that comprise the steps:
(1) communication resources logical partitioning;
(2) node loads and expands arbitration information stream;
(3) described arbitration information stream is resolved;
(4) respond according to analysis result in described (3) and service;
Wherein, described communication resources refer to realize channel or the information transmission bandwidth resource of information transmission;Described
Respond according to analysis result in described (3) and service, referring to arbitrate winning node within the execution cycle
The communication resources obtained are enjoyed exclusive communication right, utilizes the communication resources obtained to carry out message biography
Send;
Described communication resources logical partitioning refers to, described communication resources are logically divided into arbitration communication money
Source and data communication resources, wherein said arbitration communication resources are used for transmitting or carry described arbitration information stream,
Described data communication resource is for the transmission of data message;
Described expansion arbitration information stream refers to that described node adds the solicited message to described communication resources;
Arbitration information stream described in parsing, i.e. according to the number of request in arbitration information stream, it is available whether calculating has
Data channel, if also data channel can be used, then shows that this node is won in this arbitration, it is possible to holding
In line period, exclusive 1 data passage carries out data transmission.
2. referee method as claimed in claim 1, it is characterised in that described node loads and expands arbitration
In flow of information step, described node load arbitration information stream and described expansion arbitration information stream be two relative
Independent behavior, and allow behavior executed in parallel or non-executed in parallel described in the two.
3. referee method as claimed in claim 1, it is characterised in that described node is to the described communication resource
Solicited message carry out record by bit flag position, and at least include 1 bit information position.
4. referee method as claimed in claim 1, it is characterised in that described node adds described communication
During the solicited message of resource, also can increase Informationflow Control information, carry out record by increasing bit flag position.
5. referee method as claimed in claim 1, it is characterised in that after once arbitration terminates, by changing
Point of switching political loyalty expands the precedence of arbitration information stream, changes priority arbitration.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201310254247.1A CN103346908B (en) | 2013-06-24 | 2013-06-24 | A kind of communication referee method towards high-performance calculation |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201310254247.1A CN103346908B (en) | 2013-06-24 | 2013-06-24 | A kind of communication referee method towards high-performance calculation |
Publications (2)
Publication Number | Publication Date |
---|---|
CN103346908A CN103346908A (en) | 2013-10-09 |
CN103346908B true CN103346908B (en) | 2016-08-10 |
Family
ID=49281684
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201310254247.1A Active CN103346908B (en) | 2013-06-24 | 2013-06-24 | A kind of communication referee method towards high-performance calculation |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN103346908B (en) |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5982780A (en) * | 1995-12-28 | 1999-11-09 | Dynarc Ab | Resource management scheme and arrangement |
CN102104874A (en) * | 2009-12-21 | 2011-06-22 | 中兴通讯股份有限公司 | Distribution method of network transmission resource and relevant device thereof |
CN102710484A (en) * | 2012-05-26 | 2012-10-03 | 济南凌康数控技术有限公司 | Ring redundant real-time Ethernet communication method |
-
2013
- 2013-06-24 CN CN201310254247.1A patent/CN103346908B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5982780A (en) * | 1995-12-28 | 1999-11-09 | Dynarc Ab | Resource management scheme and arrangement |
CN102104874A (en) * | 2009-12-21 | 2011-06-22 | 中兴通讯股份有限公司 | Distribution method of network transmission resource and relevant device thereof |
CN102710484A (en) * | 2012-05-26 | 2012-10-03 | 济南凌康数控技术有限公司 | Ring redundant real-time Ethernet communication method |
Also Published As
Publication number | Publication date |
---|---|
CN103346908A (en) | 2013-10-09 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US8694704B2 (en) | Method and apparatus for congestion-aware routing in a computer interconnection network | |
JP6060316B2 (en) | Method and system for configuring NoC and computer-readable storage medium | |
US10476697B2 (en) | Network-on-chip, data transmission method, and first switching node | |
CN105022717A (en) | Network on chip resource arbitration method and arbitration unit of additional request number priority | |
CN109302357A (en) | A kind of on piece interconnection architecture towards deep learning reconfigurable processor | |
CN108199985B (en) | NoC arbitration method based on global node information in GPGPU | |
CN103532876A (en) | Processing method and system of data stream | |
WO2023006006A1 (en) | Roller arbitration method and circuit for on-chip data exchange | |
CN105530206B (en) | A kind of Torus network system and its working method with double access infrastructures | |
Effiong et al. | Distributed and dynamic shared-buffer router for high-performance interconnect | |
KR101994929B1 (en) | Method for operating collective communication and collective communication system using the same | |
Rad et al. | A novel arbitration mechanism for crossbar switch in wireless network-on-chip | |
CN106776014A (en) | Parallel acceleration method and system in Heterogeneous Computing | |
CN103346908B (en) | A kind of communication referee method towards high-performance calculation | |
Sun et al. | DCBuf: a high-performance wireless network-on-chip architecture with distributed wireless interconnects and centralized buffer sharing | |
CN109254844A (en) | A kind of triangle calculation method of Large Scale Graphs | |
Rad et al. | Flow control and scheduling mechanism to improve network performance in wireless NoC | |
US10768932B2 (en) | Information processing system, arithmetic processing circuit, and control method for information processing system | |
Lee et al. | Design of a feasible on-chip interconnection network for a chip multiprocessor (cmp) | |
Katta et al. | AS-Router: A novel allocation service for efficient Network-on-Chip | |
CN115955429B (en) | Network-on-chip routing method, device, system and electronic equipment | |
US20240095062A1 (en) | Offloaded task computation on network-attached co-processors | |
CN114760241B (en) | Routing method used in computing equipment of data flow architecture | |
CN117221212B (en) | Optical network on chip low congestion routing method and related equipment | |
Krishna | Enabling dedicated single-cycle connections over a shared network-on-chip |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
C14 | Grant of patent or utility model | ||
GR01 | Patent grant |