CN114697187B - Master selection method - Google Patents
Master selection method Download PDFInfo
- Publication number
- CN114697187B CN114697187B CN202210438232.XA CN202210438232A CN114697187B CN 114697187 B CN114697187 B CN 114697187B CN 202210438232 A CN202210438232 A CN 202210438232A CN 114697187 B CN114697187 B CN 114697187B
- Authority
- CN
- China
- Prior art keywords
- master
- power consumption
- value
- server
- topology
- 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
Images
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/04—Network management architectures or arrangements
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
- G06F9/5005—Allocation of resources, e.g. of the central processing unit [CPU] to service a request
- G06F9/5027—Allocation of resources, e.g. of the central processing unit [CPU] to service a request the resource being a machine, e.g. CPUs, Servers, Terminals
-
- 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
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Theoretical Computer Science (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Power Sources (AREA)
Abstract
The invention relates to a master selection method, which comprises the step S1 of acquiring topological structure information { G ] of a computing device when a server is powered on 1 ，G 2 ，…G N If G 1 ，G 2 ，…G N And connecting in sequence to form a linear topological structure, and executing the step S2: each G i Generating a request packet R i ，R i Comprising G i Id and G of i Target parameter of G i R is to be i From G in a first direction of a linear topology i‑1 In turn to G 1 ，G i R is to be i From G in a second direction of the linear topology i+1 In turn to G N (ii) a S3, obtaining each G j Received R emitted in a first direction along a linear topology i Number M of i1 And R emitted in a second direction of the linear topology i Number M of i2 If M is present i1 =M i2 Or M is i1 +1=M i2 Then G will be i Determined as master. The invention improves the information interaction efficiency among a plurality of computing devices.
Description
Technical Field
The invention relates to the technical field of computers, in particular to a master selection method.
Background
In the normal operation process of AI operation devices such as a GPU, an FPGA, a DSP, and an AI accelerator in a server, power consumption needs to be maintained near Thermal Design Power (TDP for short), an existing server or a server cluster usually includes a plurality of operation devices, and how to reasonably distribute Power consumption for the plurality of operation devices is crucial in the operation process of the plurality of operation devices. In the related art, one master (master operation device) is generally selected from a plurality of operation devices, and power consumption is allocated to the plurality of operation devices by the master. However, the prior art has at least the following disadvantages: (1) In the prior art, a master is generally selected randomly, which may cause that the distance between part of computing equipment and the master is long, the delay is large, and the efficiency of power consumption distribution is affected. (2) In the prior art, in the process of allocating power consumption to a plurality of computing devices, static allocation is usually performed only based on the TDP values of the computing devices, and the real-time working state of each computing device is not fully considered, so that power consumption waste is caused, and the power consumption utilization rate is low. Therefore, how to determine a reasonable master in a plurality of computing devices and how to reasonably distribute power consumption among the plurality of computing devices become an urgent technical problem to be solved.
Disclosure of Invention
The invention aims to provide a master selection method, which is used for quickly selecting a master for a plurality of operation devices in a server when the server is powered on every time, and selecting the master with a central position by the plurality of operation devices, so that the delay of information interaction among the operation devices is minimum, and the information interaction efficiency among the operation devices is improved.
The invention provides a master selection method, which is applied to a scene of interaction of a plurality of computing devices and comprises the following steps:
step S1, when the server is powered on, obtaining topological structure information { G) of the computing equipment 1 ，G 2 ，…G N If G 1 ，G 2 ，…G N Sequentially connecting to form a linear topological structure, and executing the step S2;
step S2, each G i Generating a request packet R i ，R i Comprising G i Id and G of i Target parameter of G i R is to be i From G in a first direction of a linear topology i-1 In turn is transmitted to G 1 ，G i R is to be i From G in a second direction of the linear topology i+1 In turn is transmitted to G N ；；
Step S3, obtaining each G j Received R emitted in a first direction along a linear topology i Number M of i1 And R emitted in a second direction of the linear topology i Number M of i2 If M is present i1 = M i2 Or M is i1 +1= M i2 Then G will be i And determining that the value ranges of i and j are 1 to N when the value range is the master, wherein N is the total number of the computing equipment in the server, and i is not equal to j.
Compared with the prior art, the invention has obvious advantages and beneficial effects. By means of the technical scheme, the master selection method provided by the invention can achieve considerable technical progress and practicability, has wide industrial utilization value and at least has the following advantages:
the invention can select the central master by the plurality of arithmetic devices when the server is powered on each time, so that the delay of information interaction among the arithmetic devices is minimum, the information interaction efficiency among the plurality of arithmetic devices is improved, and the master can be determined in a self-adaptive manner based on the number of the current arithmetic devices and the connection structure when the server is powered on each time because the number of the arithmetic devices is increased or reduced and then the arithmetic devices are powered on again.
The foregoing description is only an overview of the technical solutions of the present invention, and in order to make the technical means of the present invention more clearly understood, the present invention may be implemented in accordance with the content of the description, and in order to make the above and other objects, features, and advantages of the present invention more clearly understood, the following preferred embodiments are described in detail with reference to the accompanying drawings.
Drawings
FIG. 1 is a flowchart of a master selection method for a multi-computing device according to an embodiment of the present invention;
fig. 2 is a flowchart of a power consumption allocation method based on multiple computing devices according to an embodiment of the present invention.
Detailed Description
To further illustrate the technical means and effects of the present invention adopted to achieve the predetermined objects, the following detailed description will be given to the specific implementation and effects of the master selection method according to the present invention with reference to the accompanying drawings and preferred embodiments.
The first embodiment,
An embodiment provides a master selection method, which is applied to a scene where multiple computing devices interact with each other, as shown in fig. 1, and includes:
step S1, when the server is powered on, acquiring topological structure information { G of the computing equipment 1 ，G 2 ，…G N If G 1 ，G 2 ，…G N Sequentially connecting to form a linear topological structure, and executing the step S2;
when the server is powered on, that is, when the server is started, when the computing devices in the server need to increase, decrease or adjust the topology structure, the server needs to be shut down, and the power is re-powered on after the number of the computing devices is adjusted, and after the power is powered on, the method of the embodiment can also adaptively determine the optimal master based on the current latest number of the computing devices and the topology structure of the computing devices. The server is specifically a single server or a cluster formed by a plurality of servers, a CPU in the server may be connected to one or a plurality of computing devices, all the computing devices corresponding to the server are connected according to a preset topology structure, and the topology structure may be a linear topology structure, a full-interconnection topology structure, or a ring topology structure.
Step S2, each G i Generating a request packet R i ，R i Comprising G i Id and G of i Target parameter of (1), G i R is to be i From G in a first direction of a linear topology i-1 In turn to G 1 ，G i R is to be i From G in a second direction of the linear topology i+1 In turn to G N ；
It is understood that the first direction and the second direction are opposite directions, and taking the id of the computing device as the first direction from large to small and as the second direction from small to large as an example, when i =1, G i Capable of transferring R only in the second direction i When i = n, G i Capable of transferring R only in a first direction i . The target parameters are determined according to a specific application scenario, for example, a master is selected for Power distribution among a plurality of computing devices, and then the corresponding target parameters are static Power consumption parameters, the static Power consumption parameters may specifically include peak Power (peak Power) and Thermal Design Power (short for) values of the computing devices, the peak Power of the computing devices is the maximum Power of the computing devices, the TDP value is the Power of the computing devices under normal operation, and after the operating devices in the same row normally operate, the Power may be maintained near the TDP, may reach the peak Power occasionally, but may also fall back to the TDP quickly. However, due to the influence of the load operation of the computing device and other factors, the power required by the computing device during the actual operation process is possibleAnd the power distribution is smaller than the TDP and can also be between the TDP and the peak power, so that the power distribution of the computing equipment needs to be adjusted in time based on the working state of the computing equipment, and the power waste is avoided.
Step S3, obtaining each G j Received R emitted in a first direction of a linear topology i Number M of i1 And R emitted in a second direction of the linear topology i Number M of i2 If M is present i1 = M i2 Or M is i1 +1= M i2 Then G will be i And determining that the value ranges of i and j are 1 to N, wherein N is the total number of the computing equipment in the server, and i is not equal to j.
As will be appreciated by those skilled in the art, steps S1-S3 are performed quickly after the server is powered on, allowing the master to be quickly determined. In addition, it should be noted that, in the first embodiment, determining the master is not limited to a scenario in which power consumption of multiple computing devices is allocated, and other scenarios in which the master needs to be selected are also within the protection scope of the present application, and the computing devices may specifically include an AI computing device such as a GPU, an FPGA, a DSP, and an AI accelerator.
The method of the embodiment one can quickly select the centrally-located master for the plurality of computing devices in the server when the server is powered on each time, so that the delay of information interaction among the computing devices is minimum, the information interaction efficiency among the plurality of computing devices is improved, and the master can be determined in a self-adaptive manner based on the number of the current computing devices and the connection structure each time the computing devices are powered on because the number of the computing devices is increased or reduced and then the computing devices are powered on again.
As an example, in the step S1, if G 1 ，G 2 ，…G N Forming a full-interconnection topological structure, and directly adding a preset G x Determined as master, x is a preset value, and x =1, 2, \8230, or N. It should be noted that, because any two computing devices in the full interconnect structure are connected, the delay effects of any one computing device are consistent, and as a preferred example, the value of x may be directly set to 1.
As an example, in step S1,if G is 1 ，G 2 ，…G N Forming a ring topology, then set G y And G y+1 Does not directly transmit request data packet between them, and G 1 ，G 2 ，…G N Converted into a linear topology, and then step S2 is performed, where y ranges from 1 to N, and when y = N, y +1 is set to 1. By converting the ring structure into a linear topology, it is sufficient to subsequently determine the master directly based on step S2 and step S3. It should be noted that converting the ring topology to the linear topology does not directly convert G to the linear topology y And G y+1 The physical connection between the two is disconnected, and G is set only in the process of determining the master y And G y+1 Does not directly transmit request data packet therebetween, and has the effect of temporary disconnection, G y And G y+1 Other data packets are not affected.
As an example, in the step S2, each G i Generating a request packet R i The method comprises the following steps:
step S21, each G i Generating a request packet R based on a first packet format i The first data packet format comprises a first packet header section and a first data section, the packet header section comprises a preset request identifier, and the first data section comprises G i Id and G of i The target parameter of (1).
It should be noted that the id of the computing device may not only include identification information of the computing device, but also include, for example, receiving address information of the computing device, and the like, where the receiving address information may be a target address for receiving a master parameter and may be directly used by a subsequent second data packet and/or a third data packet, but it is understood that the second data packet and/or the third data packet may also generate an id of a corresponding computing device according to a specific requirement. In addition, when each G i Receive a terminal-connected G i The transmitted request data packet is automatically transmitted to the other end one by one after the preset request identification is identified.
When the number of the computing equipment is too large, a master is directly selected, and the computing equipment far away from the master still has larger delay, so that the implementation is realizedFor example, in the step S1, if N is not less than N 1 ，N 1 Represents a first preset threshold, which is determined according to the specific application requirements and in combination with specific parameters of the computing device, and may be set to 16, for example. The method further comprises the following steps:
step S10, adding G 1 ，G 2 ，…G N Divided into S groups L 1 ，L 2 ，…L S },L s Represents the S group, S has a value ranging from 1 to S, U s Represents L s Number of arithmetic devices of medium arithmetic device, max (U) s )≤N 2 ，N 2 Representing a second predetermined threshold value, N 2 <N 1 ；
In the preferred embodiment, G is set as much as possible 1 ，G 2 ，…G N Divided into S packets equally, e.g. N =32,n 1 =16，N 2 =8, then G 1 ，G 2 ，…G N The division is into four groups of 8 arithmetic devices.
Step S20, obtaining each G i Corresponding M i1 And M i2 And G i The value range of the belonged grouping number h, h is 1 to S based on M i1 、M i2 、h、U s Obtaining G i Corresponding first intermediate parameter MA i1 And a second intermediate parameter MA i2 ：
Wherein M is obtained based on steps S2 and S3 i1 And M i2 And will not be described in detail herein.
Step S30, grouping each group L s In, satisfy MA i1 = MA i2 Or MA i1 +1= MA i2 G of (A) i Is determined as L s Corresponding second master, determining the master obtained in step S3 as the first masterA master, the first master being a master of the second master.
As an embodiment, it is specifically possible to determine the first master through step S3, and then go through the first master to each G i Sending S and U s , G i According to S and U s Can determine G i The grouping number h to which h belongs can be determined, for example, in such a way that when h satisfies M i1 > sum(U s ) Wherein s has a value range of [1, h-1 ]](ii) a And M i2 >= sum(U s ) Wherein the value range of s is [ h +1,S']When h is G i The number of the belonging packet, wherein, sum (U) s ) Represents U s Corresponding to the sum of the values. It should be noted that the above is only one algorithm for determining h, and is not limited thereto. In addition, it is also possible to directly uniquely G at the time of grouping i And setting a corresponding group number identifier, and directly determining a corresponding h value according to the group number identifier.
It should be noted that, still taking power allocation as an example, the second master first performs power allocation on all the first masters, and each first master performs power allocation on a plurality of computing devices in a group based on the allocated power consumption, thereby improving power allocation efficiency and reducing information transmission delay.
As an embodiment, the method further comprises:
s4, the master generates a master notification data packet according to a preset second data packet format, the second data packet format comprises a second header segment and a second data segment, the second header segment comprises a preset master notification identifier, and the second data segment comprises a G corresponding to the master i Id of (2).
It is understood that after each computing device receives the master notification packet, the computing devices will be transferred one by one toward the computing device in the direction away from the master, and the details are not repeated herein.
And S4, each computing device can acquire the computing device id of the master so as to report corresponding information to the master subsequently, and each request data packet comprises the target parameters of the corresponding computing device, so that the master directly follows each R i Obtained to G i The target parameter of (1).
It can be understood that if the method further passes through the operations of grouping and grading determination of the masters in steps S10 to S30, the method in step S4 sends the computing device id of the first master to each second master, and each second master sends the corresponding computing device id to the computing devices in the group, so that the implementation details are consistent with step S4, and are not described herein again.
Besides the target parameter transmission among a plurality of computing devices, other interactive information can be transmitted among the connected computing devices, and as an embodiment, the method further includes:
step S100, G i Generating an information interaction data packet according to a preset third data packet format, wherein the third data packet format comprises a third packet header section and a third data section, the third packet header section comprises an information interaction identifier, the third data section comprises an initiating terminal id, a receiving terminal id and target interaction information, and the initiating terminal id is used for storing G for generating the information interaction data packet i The receiving end id is used for storing G of the information interaction data packet i If the preamble master informs the G in the packet, it should be noted that i The id of (2) already contains address information used by the third data packet, so that the corresponding receiving end id is not added in the third data packet, and the third data packet is directly sent based on the corresponding receiving end address information;
step S200, G i And sending the information interaction data packet to a receiving end operation device.
The master according to the first embodiment may be directly applied to the power consumption allocation method based on the multiple computing devices according to the second embodiment, and may also be applied to other application scenarios requiring selection of the master.
Example II,
An embodiment two provides a power consumption allocation method based on multiple computing devices, as shown in fig. 2, including:
step C1, when the server is electrified, all G i Sending static power consumption parameters to a master, the master being for each G i Setting initial distributionPower consumption, the master being one of the plurality of computing devices corresponding to the server, G i The value range of i is 1 to N, and N is the total number of the operation equipment in the server;
wherein the initial allocation of power may be based on G i The static power consumption parameters and the total power consumption of the server are directly distributed, or default values can be directly set, and as the master can be quickly determined after power-on and the process of dynamically distributing power consumption is quickly entered, reasonable initial distribution power consumption can be set.
Note that the master satisfies all the operations G i The sum of the time delays for transmitting the information to the master is the minimum value. The master can be directly set, or the master can be adaptively selected according to the method described in the first embodiment, which is not described herein again.
Step C2, G i G is obtained at intervals of a preset first time interval i Current { U } i ,I i ,T i ,F i }，G i Based on the current { U i ,I i ,T i ,F i } and the current allocation power consumption PA i Adjusting the internal power consumption of the computing device, wherein U i Represents G i Current value of voltage, I i Represents G i Current value of current, T i Represents G i Current temperature value, F i Represents G i A current frequency value;
it will be appreciated that through U i ,I i The instantaneous power, T, can be determined i ,F i Are all reacted with G i Is proportional to the power consumption.
Step C3, G i Every preset second time interval, based on the { U ] in the historical time window before the current time i ,I i ,T i ,F i { UV } Generation i ,IV i ,TV i ,FV i And sending to the master, UV i Represents G i Current period voltage value of IV i Represents G i Periodic current value of, TV i Represents G i Periodic temperature value of (FV) i Represents G i The second time interval is greater than the first time interval;
and preferably, the historical time window is equal to the second time interval.
Step C4, presetting total power consumption of the master based on the server and each G i { UV of (1) } i ,IV i ,TV i ,FV i Each G i Generating each G of the static power consumption parameters i Current allocated power consumption PA i ’ Update PA i = PA i ’ 。
In the second embodiment, the internal power consumption of the operation device is adjusted by the fine granularity in the operation device, the dynamic power consumption distribution among the operation devices is adjusted by combining the master coarse granularity, and the power consumption distribution counteracts the internal power consumption adjustment of each G, so that the power consumption utilization rate of the operation devices is improved, the power consumption waste is avoided, and the real-time reasonable distribution of the power consumption among the operation devices is realized.
As an example, in step C2, G i Based on the current { U i ,I i ,T i ,F i And current allocation power consumption PA i Adjusting the internal power consumption of the computing device, including:
step C21, if U i *I i -PA i >PX i ，PX i Is G i Corresponding to the floating threshold of power consumption, the control reduces G i T of i And/or F i If PA i -U i *I i >PX i Then control to raise G i T of i And/or F i . The temperature adjustment can be realized by adjusting a fan or a cooling device.
It will be understood that, under normal conditions, G i The power consumption PA should be allocated at the present i But because of each G i Real-time changes in workload, etc., will cause G i Current and PA of i The difference is large, so that G can be controlled by adjusting the temperature and/or frequency in step C21 i Power consumption, but it is understood that there is a process for power consumption adjustment, and each G i The workload and other conditions can change in real time, so that dynamic allocation is needed to be performed in real time through the master based on the power consumption states of all the computing devices, the power consumption utilization rate is improved, and power consumption waste is avoided.
As an embodiment, in the step C3, the current time is based on { U ] in the historical time window before the current time i ,I i ,T i ,F i { UV } Generation i ,IV i ,TV i ,FV i And (4) the method comprises the following steps:
all the U in the historical time window before the current time i Is determined as UV i All of I i Is determined as IV i All of T i Is determined as TV i All of F i Is determined as FV i 。
As another example, in the step C3, the current time is based on { U ] in the historical time window before the current time i ,I i ,T i ,F i { UV } Generation i ,IV i ,TV i ,FV i And (4) the method comprises the following steps:
all U in the historical time window before the current time i The maximum value in (A) is determined as UV i All of I i Is determined as IV i All of T i Is determined as TV i All of F i Is determined as FV i 。
It is understood that the above-mentioned means of taking the average or maximum value are only two examples, and can also be based on { U }according to the application requirement i ,I i ,T i ,F i Take other reasonable values to generate { UV } i ,IV i ,TV i ,FV i There is no longer a single row.
As an example, the step C4 includes:
step C41, if UV i *IV i -PA i ≥PB i ，PB i Assigning a floating threshold for power consumption, then G is set i Dynamic adjustment coefficient k of i =(UV i *IV i -PA i )/ PA i If PA i -UV i *IV i ≥PB i Then set k i =-(PA i -UV i *IV i )/ UV i *IV i If UV i *IV i -PA i │<PB, then set k i =0;
Step C42, adjusting the step length lambda and G according to the current dynamic state i Obtaining the dynamic adjustment coefficient of G i Current allocated power consumption PA i ’ =PA i +k i * λ, λ satisfies
Less than or equal to 0, updating PA i = PA i ’ 。
G with margin for the current distribution power consumption can be ensured not to exceed the total power consumption through the steps C41 to C42 i The current distribution power consumption is reduced, and G can be higher i The current distribution power consumption is increased, the utilization rate of the power consumption is improved, and in addition, the equipment can run to reach the highest frequency within the allowed range of the power consumption, so that the task on the equipment can be completed more quickly, and the performance of the equipment is improved.
As an example, the first time interval is set to be in the order of milliseconds, and G is set to be within the first time interval i Is smaller than a preset temperature change threshold, e.g. the first time interval takes 1ms. The second time interval is set to the order of seconds, for example the second time interval is set to 1s.
It should be noted that the same technical details that have already been described in the first embodiment are not described again in the second embodiment, and the technical details related to the first embodiment and the second embodiment may be used in combination.
It should be noted that some of the exemplary embodiments in the first and second embodiments are described as processes or methods depicted as flowcharts. Although a flowchart may describe the steps as a sequential process, many of the steps can be performed in parallel, concurrently or simultaneously. Additionally, the order of many of the steps may be rearranged. A process may be terminated when its operations are completed, but may have additional steps not included in the figure. A process may correspond to a method, a function, a procedure, a subroutine, a subprogram, etc.
An embodiment of the present invention further provides an electronic device, including: at least one processor; and a memory communicatively coupled to the at least one processor; wherein the memory stores instructions executable by the at least one processor and configured to perform a method according to an embodiment of the invention.
The embodiment of the invention also provides a computer-readable storage medium, and the computer instructions are used for executing the method of the embodiment of the invention.
Although the present invention has been described with reference to the preferred embodiments, it will be understood by those skilled in the art that various changes may be made and equivalents may be substituted for elements thereof without departing from the scope of the present invention.
Claims (8)
1. A master selection method is applied to a scene of interaction of a plurality of arithmetic devices, and comprises the following steps:
step S1, when the server is powered on, acquiring topological structure information { G of the computing equipment 1 ，G 2 ，…G N If G 1 ，G 2 ，…G N Sequentially connecting to form a linear topological structure, and executing the step S2;
step S2, each G i Generating a request packet R i ，R i Comprising G i Id and G of i Target parameter of (1), G i R is to be i From G in a first direction of a linear topology i-1 In turn is transmitted to G 1 ，G i R is to be i From G in a second direction of the linear topology i+1 In turn is transmitted to G N ；
Step S3, obtaining each G j Received R emitted in a first direction of a linear topology i Number M of i1 And R emitted in a second direction of the linear topology i Number M of i2 If M is present i1 ＝M i2 Or M is i1 +1＝M i2 Then G will be i And determining that the value ranges of i and j are 1 to N, wherein N is the total number of the computing equipment in the server, and i is not equal to j.
2. The method of claim 1,
in the step S1, if G 1 ，G 2 ，…G N Forming a full-interconnection topological structure, and directly connecting the preset G x Determined as master, x is a preset value, and x =1, 2, \8230, or N.
3. The method of claim 2,
the value of x is 1.
4. The method of claim 1,
in the step S1, if G 1 ，G 2 ，…G N Forming a ring topology, then set G y And G y+1 Does not directly transmit request data packet between them, and G 1 ，G 2 ，…G N Converted into a linear topology, and then step S2 is performed, where y ranges from 1 to N, and when y = N, y +1 is set to 1.
5. The method of claim 1,
in the step S1, if N is more than or equal to N 1 ，N 1 Which is indicative of a first pre-set threshold value,the method further comprises the following steps:
step S10, adding G 1 ，G 2 ，…G N Divided into S packets { L } 1 ，L 2 ，…L S },L s Represents the S group, S has a value ranging from 1 to S, U s Represents L s Number of arithmetic devices, max (U) s )≤N 2 ，N 2 Representing a second predetermined threshold value, N 2 <N 1 ；
Step S20, obtaining each G i Corresponding M i1 And M i2 And G i The value range of the belonged grouping number h, h is 1 to S based on M i1 、M i2 、h、U s Obtaining G i Corresponding first intermediate parameter MA i1 And a second intermediate parameter MA i2 ：
Step S30, grouping each group L s In, satisfy MA i1 ＝MA i2 Or MA i1 +1＝MA i2 G of (A) i Is determined as L s And the corresponding second master determines the master obtained in the step S3 as the first master, wherein the first master is the main master of the second master.
6. The method of claim 5,
the first master first allocates power consumption to all the second masters, and each second master allocates power consumption to a plurality of arithmetic devices in a group based on the allocated power consumption.
7. The method of claim 5,
in the step S10, G is added 1 ，G 2 ，…G N Are equally divided into S packets.
8. The method of claim 1,
the operation equipment is a GPU, an FPGA, a DSP or an AI accelerator.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210438232.XA CN114697187B (en) | 2022-04-25 | 2022-04-25 | Master selection method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210438232.XA CN114697187B (en) | 2022-04-25 | 2022-04-25 | Master selection method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114697187A CN114697187A (en) | 2022-07-01 |
| CN114697187B true CN114697187B (en) | 2022-12-02 |
Family
ID=82144685
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210438232.XA Active CN114697187B (en) | 2022-04-25 | 2022-04-25 | Master selection method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114697187B (en) |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109408257A (en) * | 2018-11-09 | 2019-03-01 | 北京灵汐科技有限公司 | Data transmission method, device and electronic equipment for network-on-chip NOC |
| CN111817953A (en) * | 2020-06-19 | 2020-10-23 | 新华三技术有限公司成都分公司 | Method and device for electing master equipment based on Virtual Router Redundancy Protocol (VRRP) |
| WO2022001086A1 (en) * | 2020-06-29 | 2022-01-06 | 苏州浪潮智能科技有限公司 | Efficient gpu resource allocation optimization method and system |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10609118B2 (en) * | 2017-03-14 | 2020-03-31 | International Business Machines Corporation | Adaptive communication control device |
-
2022
- 2022-04-25 CN CN202210438232.XA patent/CN114697187B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109408257A (en) * | 2018-11-09 | 2019-03-01 | 北京灵汐科技有限公司 | Data transmission method, device and electronic equipment for network-on-chip NOC |
| CN111817953A (en) * | 2020-06-19 | 2020-10-23 | 新华三技术有限公司成都分公司 | Method and device for electing master equipment based on Virtual Router Redundancy Protocol (VRRP) |
| WO2022001086A1 (en) * | 2020-06-29 | 2022-01-06 | 苏州浪潮智能科技有限公司 | Efficient gpu resource allocation optimization method and system |
Non-Patent Citations (1)
| Title |
|---|
| 基于片上网络的路由模型研究;张浩 等;《航空计算技术》;20080331;第38卷(第2期);第120-122页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114697187A (en) | 2022-07-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109617826B (en) | Storm dynamic load balancing method based on cuckoo search | |
| US7982336B2 (en) | Power sharing with stackable switches | |
| WO2022111453A1 (en) | Task processing method and apparatus, task allocation method, and electronic device and medium | |
| CN113784373B (en) | Combined optimization method and system for time delay and frequency spectrum occupation in cloud edge cooperative network | |
| WO2019134197A1 (en) | Method and system for selecting minimum load router based on naive bayes classifier | |
| CN104243405A (en) | Request processing method, device and system | |
| CN113364850A (en) | Software-defined cloud-edge collaborative network energy consumption optimization method and system | |
| CN111245924A (en) | Load balancing method and device and computer storage medium | |
| CN113986562A (en) | Resource scheduling strategy generation method and device and terminal equipment | |
| CN113382074A (en) | Micro-service load balancing optimization method based on dynamic feedback | |
| CN114697187B (en) | Master selection method | |
| CN110545315A (en) | heartbeat interval adjusting method based on data block quantity change and bandwidth change | |
| CN114546666B (en) | Power consumption distribution method based on multiple computing devices | |
| CN114546095B (en) | Master selection method based on multiple computing devices | |
| CN111240824A (en) | CPU resource scheduling method and electronic equipment | |
| CN111858029B (en) | Storm cluster load balancing method and system based on discrete particle swarm | |
| Bestavros et al. | Probabilistic job scheduling for distributed real-time applications | |
| CN112437449A (en) | Joint resource allocation method and area organizer | |
| CN115168017B (en) | Task scheduling cloud platform and task scheduling method thereof | |
| CN107729141B (en) | Service distribution method, device and server | |
| Billard et al. | Dynamic scope of control in decentralized job scheduling | |
| CN110636104B (en) | Resource request method, electronic device and storage medium | |
| Zhu et al. | Load balancing algorithm for web server based on weighted minimal connections | |
| CN114546630A (en) | Task processing method and distribution method, device, electronic equipment and medium | |
| CN112732442B (en) | Distributed model for edge computing load balancing and solving method thereof |
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 |