CN111400026B - Distributed load balancing method based on master-slave backup technology - Google Patents
Distributed load balancing method based on master-slave backup technology Download PDFInfo
- Publication number
- CN111400026B CN111400026B CN201911119106.2A CN201911119106A CN111400026B CN 111400026 B CN111400026 B CN 111400026B CN 201911119106 A CN201911119106 A CN 201911119106A CN 111400026 B CN111400026 B CN 111400026B
- Authority
- CN
- China
- Prior art keywords
- task
- node
- tasks
- nodes
- master
- 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
-
- 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]
- G06F9/5083—Techniques for rebalancing the load in a distributed system
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F11/00—Error detection; Error correction; Monitoring
- G06F11/07—Responding to the occurrence of a fault, e.g. fault tolerance
- G06F11/14—Error detection or correction of the data by redundancy in operation
- G06F11/1402—Saving, restoring, recovering or retrying
- G06F11/1446—Point-in-time backing up or restoration of persistent data
- G06F11/1448—Management of the data involved in backup or backup restore
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/11—Complex mathematical operations for solving equations, e.g. nonlinear equations, general mathematical optimization problems
- G06F17/12—Simultaneous equations, e.g. systems of linear equations
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Mathematical Analysis (AREA)
- Computational Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Optimization (AREA)
- Software Systems (AREA)
- Pure & Applied Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Quality & Reliability (AREA)
- Operations Research (AREA)
- Algebra (AREA)
- Databases & Information Systems (AREA)
- Multi Processors (AREA)
Abstract
The invention discloses a distributed load balancing method of a master-slave backup technology, which comprises the following steps: s1, dividing a node set; s2, dividing a task set; s3, distributing subtasks; s4, recording the execution time; s5, calculating execution efficiency; s6, solving a distribution scheme; by modeling the task allocation problem as a linear programming problem, and by analyzing the abstract model, the solution of the task allocation feasible domain is realized. Meanwhile, in combination with practical problems, a task allocation and dynamic adjustment method with backup is provided, and the method aims to realize task allocation according to efficiency on a distributed cluster with backup and reduce apparent performance relevance among nodes as much as possible. The method and the device can avoid increasing the extra burden of the backup node as much as possible when the task amount of the main node is increased, and further can relieve the influence of performance difference between the nodes on the whole task running time.
Description
Technical Field
The invention belongs to the field of computer communication, and particularly relates to a distributed load balancing method based on a master-slave backup technology.
Background
In distributed systems, the key way to tolerate a process failure is to put multiple identical processes into a group, and when a message is sent to the group itself for processing, all members accept and process it. In this way, if one process in the group fails, some other process may take over for it. Under the condition of needing fault tolerance, a process copying method is usually used, and when a main process crashes, a backup process replaces the task of the current main process; in the case of needing to improve the performance, the main process is usually copied by using copy and cache extension and redundancy coding to form a process group, tasks can be recovered more quickly due to the existence of redundancy (task redundancy part based on MDS codes is recovered quickly), and the communication quantity required by the tasks can be reduced (communication load based on CDC coding is reduced).
In a conventional process backup scheme, a main process running on a main node is completely copied to form a backup process, and the backup process is placed on a backup node. Because the nodes have a full backup relationship, strong consistency needs to be ensured between the main node and the backup nodes, and the task quantity of the main node is increased while the task quantity of the backup nodes is correspondingly increased. When the performance of the master node and the backup node in the group is different, increasing the task load of the master node may cause a non-negligible running time difference between the master node and the backup node, which affects the real-time performance of backup.
Disclosure of Invention
The invention aims to: in order to model the task allocation problem into a linear programming problem, the solution of a task allocation feasible domain is realized through the analysis of an abstract model, and meanwhile, a task allocation and dynamic adjustment method with backup is provided by combining with an actual problem.
The technical scheme is as follows: the invention provides a distributed load balancing method based on a master-slave backup technology, which comprises the following specific steps:
(1) Dividing a node set according to the constructed node distributed cluster;
(2) Running a task for setting a backup level on a cluster, and dividing all tasks in the same batch into task sets;
(3) Distributing tasks contained in the task set to the corresponding node set according to the node set and the task set obtained in the steps (1) and (2);
(4) After all the nodes finish the received tasks, collecting and recording the execution time;
(5) Acquiring the calculation efficiency of each node, and normalizing the calculation execution efficiency;
(6) And establishing an equation set through known constants, converting the deformation into a linear programming problem, and further solving the distribution scheme.
Further, the method for partitioning a node set in step (1) includes:
(1.2) arranging and combining n nodes, and taking any r nodes to form a group of subsetsAll possible combinationsComposition setCollectionIs provided withThe number of the elements is one,initialization
Further, the method for dividing the task set in step (2) includes:
(2.1) running F tasks with equal task quantity on the cluster constructed in the step (1), and setting a backup level as r;
(2.2) dividing the whole task set F into small batches, performing task allocation and operation statistics once in each batch, recording the system operation time period as t, and recording the batches as F j Each batch F j The calculated time of (d) is recorded as Δ t; each batch F j Divided into smaller sets of small tasks, denoted The size of (A) is recorded as
(2.3) all tasks F of the same batch j Division intoA set of tasks, at time t, in proportionBatch dividing task F j Is composed ofA task setCollectionAndlike, setAlso haveAnd (4) each element.
Further, the method for distributing subtasks in step (3) includes:
(3.1) according to the steps (1) and (2), two genes are obtainedSet of individual elementsAndtaking elements from two collections at a timeAndwhereinIs a set of tasks that is to be executed,is a set of nodes;
(3.2) sequentially subjectingThe contained task is sent toRepeating the above process on the represented nodes until each task setAll the tasks in (1) are sent to the correspondingOn all nodes in the node; the number of the tasks to be processed existing on each node is the sameEach task set is copied and sent to r different nodes.
Further, the step (4) of recording the execution time process further includes:
and (4.1) waiting for all the nodes to complete the received tasks, and executing redundancy check or other tasks. Collecting each node N q Is recorded as Δ t q ;
(4.2) recording the current time as t, node N q The current computational efficiency of is noted as λ q (t) sectionPoint N q Task set owned byThe size of (A) is recorded as
And (4.3) calculating by adopting the formula (1).
Each node independently calculates the calculation efficiency lambda of the node q (t) and sends it to all other nodes.
Further, the method for calculating the execution efficiency in the step (5) comprises the following steps:
(5.1) obtaining all lambda of each node q (t) using equation (2) for λ q (t) performing normalization;
e(t)=(e 0 (t),e 1 (t),…,e n-1 (t));
(5.2) recording the distribution matrix asLet A be the coefficient matrix and,as variables, e (t) is a constant term, and a non-homogeneous linear equation system (3) can be obtained.
Wherein, the row mark of the distribution matrix represents the node subscript, the column mark represents the task set subscript, and the element a q,i =1 representing node q owning task set i, element a q,i =0 representing node q not owning task set i
Further, the solving of the allocation scheme in the step (6) specifically includes:
(7.1) when the number of rows is equal to the number of columns,the system of equations has a unique solution
(7.2) when the number of rows in the system of equations is greater than the number of columns,the equation system has no unique solution and can be transformed into a linear programming problem, and the process is as follows:
The above equation set is rewritten as (equation 4):
solving to obtain:
if it isIs greater than 0, is takenAs an argument, it is converted into a linear programming problem, as in equation (6), whereIs a vectorThe ith element of (c):
(7.2.2) vectorEach obtained in ZhongdeAn element Corresponding to the ratio of a task set to the total number of tasks, vectorThe corresponding set of values can be used as a set of allocation schemes. Divide each batch of tasks F by a ratio using the allocation scheme j And (5) letting t = t +1, entering the next time step, executing the step (2), and repeating the task allocation and the performance estimation calculation until all the tasks are executed.
Has the beneficial effects that: compared with the prior art, the invention has the remarkable advantages that: (1) The method can realize task allocation according to efficiency on the backup distributed cluster, and simultaneously reduce apparent performance relevance among nodes as much as possible; (2) The method and the device can avoid increasing the extra burden of the backup node as much as possible when the task amount of the main node is increased, and further can relieve the influence of performance difference between the nodes on the whole task running time.
Drawings
Fig. 1 is a flowchart of a distributed load balancing method based on a master-slave backup technology;
FIG. 2 may be a domain sample.
The specific implementation mode is as follows:
the following description will explain embodiments of the present invention by referring to the drawings.
Example 1 is specifically described as figure 1: setting a distributed cluster with 3 nodes If it is to be atThe above 4 times of loop calculation are performed, and tasks with 120 equal task quantities are performed at a time (the tasks have no relevance, can be performed out of order and do not influence each other). The number r =2 of redundant backups of each task on the present cluster is set (i.e., each task is to be executed twice at the same time). Embodiments of the method according to the above steps will be described below.
Step 1: arranging and combining the 3 nodes to obtain a combined result Initializing per-batch task set size
And 2, step: according to the size of each batch of task setsDividing the tasks in each batch of circulation, and dividing a batch of 120 subtasks into three setsAnd is
And step 3: according to steps 1 and 2, two sets of 3 elements are obtainedAndtaking elements from two collections at a timeAndwhereinIs a set of tasks that is to be executed,is a collection of a set of nodes. In turn willThe contained task is sent toOn the node of the representation, i.e. each subsetAll nodes in the network receive the same task setThat is to say, theSend to node { N 0 ,N 1 Will beTo a node N 0 ,N 2 Will beTo a node N 1 ,N 2 }. At this time, the number of the tasks to be processed existing on each node is μ =2 in total, each task set is copied and sent to 2 different nodes, and then the step 4 is performed.
And 4, step 4: and waiting for all the nodes to complete the received tasks. Each node N q And independently recording the computing efficiency of all the subtasks and sending the computing efficiency to all other nodes in the cluster. With node N 0 For example, the total running time is recorded as Δ t 0 At the current moment t, the node N is connected 0 Is provided withTask set of { P } 1 ,P 2 The size of the symbol is recorded asCalculating lambda 0 (t)。
And sends λ 0 (t) to the node N 1 ,N 2 . Proceed to step 5.
And 5: at each node N q The following operations are performed: obtaining all λ from other nodes q (t) for λ q (t) normalizing to obtain
Definition ofAnd e (t) is of the form: e(t)=(e 0 (t),e 1 (t),e 2 (t)). A non-homogeneous system of linear equations can be obtained:
wherein:
step 6: solving a non-homogeneous system of linear equationsAnd as a task allocation scheme of the next iteration, continuing to execute the step 2 until the iteration of all batches is completed.
Suppose that in an iteration, node N 0 ,N 1 ,N 2 Are respectively delta t 0 =16.7secs,Δt 1 =11.1secs,Δt 2 =14.3secs. The calculation efficiency is obtained as lambda 0 (t)=0.0399,λ 1 (t)=0.0600,λ 2 (t) =0.0466, and after normalization e can be obtained 0 (t)=0.545,e 1 (t)=0.818,e 2 (t) =0.636 the equation representation of the above-described heterogeneous linear system of equations can be written as:
it can be solved that:
then the next time a task is divided, the size of each of its subtask sets The total number of tasks at each node becomes 66,98,76, which can be considered to achieve approximately maximum resource utilization, corresponding to its computed performance estimate of 0.0399,0.0600, 0.0466.
Example 2: setting a distributed cluster with 4 nodesIf it is to be atExecute 4 loop calculations, 120 each timeTasks with equal task amount (tasks have no relevance, can be executed out of order and do not influence each other). The number r =2 of redundant backups of each task on the present cluster is set (i.e., each task is to be executed twice at the same time). Embodiments of the method according to the above steps will be described below.
Step 1: 4 nodes are arranged and combined, and the total number isSpecies combination scheme, denoted D 1 ,D 2 ,…,D 6 。D 1 ={N 1 ,N 2 },D 2 ={N 1 ,N 3 },…,D 6 ={N 3 ,N 4 }. The combined results are recorded as Initialization
Step 2: according to the size of each batch of task setsDividing the tasks of each batch of circulation, then a batch of 120 sub-tasks will be divided into 6 setsAnd is
And step 3: according to steps 1 and 2, two sets of 6 elements are obtainedAndtaking elements from two collections at a timeAndwhereinIs a set of tasks that is to be executed,is a collection of nodes. In turn willThe contained task is sent toOn the node of the representation, i.e. each subsetAll nodes in the network receive the same task setThat is to say thatTo a node N 0 ,N 1 Will beTo a node N 0 ,N 2 Will beSend to node { N 0 ,N 3 And so on. At this time, the number of tasks to be processed existing on each node is μ =3 in total, and each task set is copied and sent to 2 different nodes, and then the process proceeds to step 4.
And 4, step 4: waiting for all nodes to complete their received tasks. Each node N q And independently recording the computing efficiency of all the subtasks and sending the computing efficiency to all other nodes in the cluster. With node N 0 For example, the total operating time is denoted as Δ t 0 When the current time is t, connecting the node N 0 Task set P owned by 1 ,P 2 ,P 3 The size of the symbol is recorded asCalculating lambda 0 (t)。
And transmit lambda 0 (t) to node N 1 ,N 2 ,N 3 . Proceed to step 5.
And 5: at each node N q The following operations are performed: obtaining all λ from other nodes q (t), for λ q (t) normalizing to obtain
a non-homogeneous system of linear equations can be obtained:
wherein:
step 6: the sparse matrix of the non-homogeneous linear equation set is not a non-singular matrix, the non-homogeneous linear equation set cannot be solved, and the non-homogeneous linear equation set can be transformed into a linear programming problem, and the process is as follows:
firstly, the matrix A can be written into a form of splicing a square matrix and a common matrix, and the left half part of the matrix A is called A l And the right half is A r Namely:
The above equation set can be rewritten as (4):
it can be solved that:
order toIs greater than 0, is takenAs an argument, it can be converted into a linear programming problem, as in equation (6).
To obtainThen, the formula (6) is substituted to solveTo obtain(Vector)Get each element Corresponding to the proportion of a task set to the total number of tasks, then vectorThe corresponding set of values can be used as a set of allocation schemes. Divide each batch of tasks F by a ratio using the allocation scheme j And (3) letting t = t +1, entering the next time step, executing the step 2, and repeating the task allocation and the performance estimation calculation until all the tasks are executed.
Suppose that in an iteration, node N 0 ,N 1 ,N 2 ,N 3 Are respectively delta t 0 =16.7secs,Δt 1 =11.1secs,Δt 2 =14.3secs,Δt 3 =11.1secs. The calculation efficiency can be obtained as follows:
λ 0 (t)=0.0299,λ 1 (t)=0.0450,λ 2 (t)=0.0349,λ 3 (t)=0.0450
after normalization, one can obtain:
e 0 (t)=0.3862,e 1 (t)=0.5814,e 2 (t)=0.4510,e 3 (t)=0.5814
the equation representation of the above-described system of non-homogeneous linear equations can be written as:
namely:
coefficient matrix:
inverting it to obtain:
order:
the following can be obtained:
the constraint has a feasible region in the first quadrant, as shown in FIG. 2.
In the above feasible domain:
taking a set of feasible solutions:
Then the next time a task is divided, the size of each of its subtask sets Therefore each timeThe total number of tasks on each node becomes 46,70,54,70, which can be considered to achieve approximately maximum resource utilization, corresponding to its computed performance estimate of 0.0299,0.0450,0.039, 0.0450.
Claims (5)
1. A distributed load balancing method based on a master-slave backup technology is characterized by comprising the following steps:
(1) Dividing a node set according to the constructed node distributed cluster;
(2) Running a task for setting a backup level on a cluster, and dividing all tasks in the same batch into task sets;
(3) Distributing tasks contained in the task set to the corresponding node set according to the node set and the task set obtained in the steps (1) and (2);
(4) After all the nodes finish the received tasks, collecting and recording the execution time;
(5) Acquiring the calculation efficiency of each node, and normalizing the calculation execution efficiency;
(6) Establishing an equation set through a known constant, converting deformation into a linear programming problem, and further solving a distribution scheme;
the method for partitioning a node set in step (1) comprises the following steps:
(1.2) arranging and combining n nodes, and taking r nodes to form a group of subsets All possible combinationsComposition setCollectionIs provided withThe number of the elements is one,initialization
The solving of the allocation scheme in the step (6) specifically comprises:
(6.1) when the number of rows is equal to the number of columns,the system of equations has a unique solution
(6.2) when the number of rows in the system of equations is greater than the number of columns,the equation set has no unique solution and can be transformed into a linear programming problem, and the process is as follows:
The above equation set is rewritten as (equation 4):
solving to obtain:
if it isIs greater than 0, is takenAs an argument, it is converted into a linear programming problem, as in equation (6), whereIs a vectorThe ith element of (c):
(6.2.2) vectorsGet each element Corresponding to the ratio of a task set to the total number of tasks, vectorThe corresponding set of values can be used as a set of allocation schemes by which to divide the tasks F of each batch in proportion j And (3) enabling t = t +1, entering the next time step, executing the step (2), and repeating task allocation and performance estimation calculation until all tasks are executed.
2. The distributed load balancing method based on the master-slave backup technology as claimed in claim 1, wherein the task set partitioning method in step (2) comprises:
(2.1) running F tasks with equal task quantity on the cluster constructed in the step (1), and setting a backup level as r;
(2.2) dividing the whole task set F into small batches, performing task allocation and operation statistics once in each batch, recording the system operation time period as t, and recording each batchNext task F j The calculated time of (d) is recorded as Δ t; task F for each batch i Divided into smaller sets of small tasks, denoted The size of (A) is recorded as
3. The distributed load balancing method based on the master-slave backup technology as claimed in claim 2, wherein the method for distributing the subtasks in step (3) comprises:
(3.1) obtaining twoSet of individual elementsAndtaking elements from two sets at a timeAndwhereinIs a set of tasks that is to be executed,is a set of nodes;
(3.2) sequentially subjectingThe included task is sent toOn the nodes of the representation up to each task setAll tasks in (1) are sent to correspondingAll nodes in the network; the number of the tasks to be processed existing on each node is the sameEach task set is copied and sent to r different nodes.
4. The distributed load balancing method based on the master-slave backup technology as claimed in claim 3, wherein the step (4) of recording the execution time further comprises:
(4.1) waiting for all the nodes to complete the tasks received by the nodes, executing redundancy check or other tasks, and collecting each node N q Is recorded as Δ t q ;
(4.2) recording the current time as t, node N q The current computational efficiency of is noted as λ q (t), node N q Task set owned byThe size of (A) is recorded as
(4.3) calculating by using the formula (1):
each node independently calculates the calculation efficiency lambda of the node q (t) and sends it to all other nodes.
5. The distributed load balancing method based on the master-slave backup technology as claimed in claim 4, wherein the method for calculating the execution efficiency in step (5) comprises:
(5.1) obtaining all lambda of each node q (t) using equation (2) for λ q (t) normalization is carried out;
(5.2) recording the distribution matrix as A,let A be the coefficient matrix and,as variables, e (t) is a constant term, a non-homogeneous linear equation system (3) can be obtained:
wherein, the row mark of the distribution matrix represents the node subscript, the column mark represents the task set subscript, and the element a q,i =1 representing node q owning task set i, element a q,i =0 represents that node q does not own task set i.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201911119106.2A CN111400026B (en) | 2019-11-15 | 2019-11-15 | Distributed load balancing method based on master-slave backup technology |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201911119106.2A CN111400026B (en) | 2019-11-15 | 2019-11-15 | Distributed load balancing method based on master-slave backup technology |
Publications (2)
Publication Number | Publication Date |
---|---|
CN111400026A CN111400026A (en) | 2020-07-10 |
CN111400026B true CN111400026B (en) | 2023-02-28 |
Family
ID=71433924
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201911119106.2A Active CN111400026B (en) | 2019-11-15 | 2019-11-15 | Distributed load balancing method based on master-slave backup technology |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111400026B (en) |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111858721B (en) * | 2020-08-03 | 2023-07-21 | 南京大学 | Distributed computing method based on priority coding |
CN113505021B (en) * | 2021-05-26 | 2023-07-18 | 南京大学 | Fault tolerance method and system based on multi-master-node master-slave distributed architecture |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103176850A (en) * | 2013-04-10 | 2013-06-26 | 国家电网公司 | Electric system network cluster task allocation method based on load balancing |
CN104283948A (en) * | 2014-09-26 | 2015-01-14 | 东软集团股份有限公司 | Server cluster system and load balancing implementation method thereof |
CN105302649A (en) * | 2015-12-03 | 2016-02-03 | 中国联合网络通信集团有限公司 | Disaster recovery backup method and system |
US20160179642A1 (en) * | 2014-12-19 | 2016-06-23 | Futurewei Technologies, Inc. | Replicated database distribution for workload balancing after cluster reconfiguration |
CN110190991A (en) * | 2019-05-21 | 2019-08-30 | 华中科技大学 | A kind of fault-tolerance approach of distributed stream processing system under more application scenarios |
-
2019
- 2019-11-15 CN CN201911119106.2A patent/CN111400026B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103176850A (en) * | 2013-04-10 | 2013-06-26 | 国家电网公司 | Electric system network cluster task allocation method based on load balancing |
CN104283948A (en) * | 2014-09-26 | 2015-01-14 | 东软集团股份有限公司 | Server cluster system and load balancing implementation method thereof |
US20160179642A1 (en) * | 2014-12-19 | 2016-06-23 | Futurewei Technologies, Inc. | Replicated database distribution for workload balancing after cluster reconfiguration |
CN105302649A (en) * | 2015-12-03 | 2016-02-03 | 中国联合网络通信集团有限公司 | Disaster recovery backup method and system |
CN110190991A (en) * | 2019-05-21 | 2019-08-30 | 华中科技大学 | A kind of fault-tolerance approach of distributed stream processing system under more application scenarios |
Non-Patent Citations (1)
Title |
---|
SDN期末作业;bokerr;《https://www.cnblogs.com/bokers/p/8343502.html》;20180124;第1-8页 * |
Also Published As
Publication number | Publication date |
---|---|
CN111400026A (en) | 2020-07-10 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111400026B (en) | Distributed load balancing method based on master-slave backup technology | |
CN102567080B (en) | Virtual machine position selection system facing load balance in cloud computation environment | |
CN105740084A (en) | Cloud computing system reliability modeling method considering common cause fault | |
Zhang et al. | MrHeter: improving MapReduce performance in heterogeneous environments | |
Xu et al. | A (DP) $^ 2 $2 SGD: Asynchronous Decentralized Parallel Stochastic Gradient Descent With Differential Privacy | |
CN110362780B (en) | Large data tensor canonical decomposition calculation method based on Shenwei many-core processor | |
CN105373517A (en) | Spark-based distributed matrix inversion parallel operation method | |
Wang et al. | Heterogeneity-aware gradient coding for straggler tolerance | |
CN110929884A (en) | Classification method and device for distributed machine learning optimization based on column division | |
CN114201287B (en) | Method for cooperatively processing data based on CPU + GPU heterogeneous platform | |
CN112948123B (en) | Spark-based grid hydrological model distributed computing method | |
CN111428192A (en) | Method and system for optimizing high performance computational architecture sparse matrix vector multiplication | |
CN115718986B (en) | Multi-core parallel time domain simulation method based on distributed memory architecture | |
CN112612601A (en) | Intelligent model training method and system for distributed image recognition | |
Mercan et al. | Computing sequence covering arrays using unified combinatorial interaction testing | |
CN111258730A (en) | Task scheduling method based on competition conditions | |
CN111722923A (en) | Heterogeneous resource calling method and device and computer readable storage medium | |
CN113342313B (en) | Method for asynchronously updating linear classification model parameters in Spark MLlib based on parameter server | |
CN111858721B (en) | Distributed computing method based on priority coding | |
Umesha et al. | Optimal design of truss structures using parallel computing | |
JPWO2005029352A1 (en) | Parallel computing method and apparatus | |
CN110021339B (en) | Cluster parallel computing acceleration method based on protein folding calculation protein structure | |
Huang et al. | Speedup and synchronisation overhead analysis of Gauss-Seidel type algorithms on a Sequent balance machine | |
Bošanský et al. | Parallel Approach To Solve Of The Direct Solution Of Large Sparse Systems Of Linear Equations | |
Sharadapriyadarshini et al. | Formulations and heuristics for scheduling in a buffer-constrained flowshop and flowline-based manufacturing cell with different buffer-space requirements for jobs: Part 2 |
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 |