RU2815189C1 - Method for time synchronization of operation of massively parallel computing system with distributed memory - Google Patents

Abstract

FIELD: computer engineering.

SUBSTANCE: invention relates to computer engineering for processing digital data. Technical result is achieved by determining number of fragment, having minimum value of cardinalities of operators sets, including the time tier operator in the fragment and calculating the fragmentation complexity for the current fragment; calculating total complexity of task fragmentation; clustering fragments; evaluating the current state of the resource of the processors with cache memory and the communication network; setting a one-to-one correspondence between clusters of fragments and a subset of currently free processors with cache memory of computing nodes; evaluation of a specific type of processors with cache memory; generating a plurality of fragments for superscalar processors and super-long command line processors; evaluation of a specific type of communication network topology; estimating maximum time spent on messaging; development of text specifications of threads of parallel program with time parameterization, taking into account the time of transmitting/receiving data into the transmitting and receiving side buffers, and adding built-in time synchronization means to the code for each processor with cache memory of computing nodes; evaluation of correctness of results of development of parallel programs with time parameterisation; compiling the generated parallel code with time parameterisation.

EFFECT: high efficiency of digital information processing (reduced time for solving a problem) of a computer system.

1 cl, 2 dwg

Description

The invention relates to the field of digital data processing using electronic devices, namely to methods for time synchronization of a computer system intended for massively parallel computing systems with distributed memory (Massive Parallel Processing (MPP)) and can be used to reduce the processing time of digital data MRR computing systems.

There is a known method for constructing a program, which consists in defining marked loops in the source code of an assembler program and classifying them into several predefined types, aligning the start addresses of the marked loops, if required for a loop of a given type, by adding assembly instructions and storing the assembler source code in memory , build a modified assembly code for the destination device by compiling and linking [1].

There is a known method for creating a parallel program with time parameterization of multiprocessor computer systems with the same access to memory, which consists in the fact that to implement the method, the control device, the instruction fetch unit and the arithmetic-logical unit perform the following operations: receives an ordered selection of tokens of the original sequential program and forms for tokens of its descriptor; receives an ordered selection of lexeme descriptors from the set of descriptors generated at the previous stage; forms new program specification structures with detail down to operations/functions; forms new program specification structures with detail down to fragments; checks the equivalence of the text specification of the task program and its representation by new specification structures; calculates priority values for operators of the new specification; generates sets of operators-candidates for the start of execution at a moment in time; assigns operators for implementation on a free processor; develops text specifications of parallel program threads with time parameterization; evaluates the correctness of the results of developing parallel programs with time parameterization; compiles the generated parallel code with time parameterization [2].

The disadvantage of these methods is the lack of consideration of the specific architecture of massively parallel computing systems with distributed memory and the lack of consideration of the time parameter.

One of the possible ways to increase the efficiency of information processing is to organize the distribution of program data between computing nodes. However, when distributing program data between computing nodes, problems arise with the synchronization of computing nodes, which leads to a decrease in the efficiency of digital information processing, which must be taken into account when choosing the start time of execution of fragments/operators during parallel execution of the original program.

The purpose of the invention is to increase the efficiency of digital information processing (reducing the time to solve a problem) due to time synchronization of the operation of a massively parallel computing system with distributed memory.

This goal is achieved by a method of time synchronization of the operation of a massively parallel computing system with distributed memory, which consists in performing the following procedures:

1. Ordered selection of tokens of the original sequential program, determination of their belonging to one or another class of tokens of a high-level programming language, verification of the token's membership in the generated (at the considered point in time) set of tokens, inclusion of the token in the generated set of tokens of the corresponding type (if it is not present in composition of the set), the formation of its descriptor for the lexeme, which specifies the numeric encoding of the necessary attributes.

2. Ordered selection of lexeme descriptors from the set of descriptors formed at the previous stage, determining the high-level programming language construction corresponding to the descriptor under consideration, forming a postfix specification for this construction based on the method of generating reverse Polish notation in accordance with Dijkstra’s algorithm based on the use of the “stack mechanism” with priorities, which allows you to change the order of symbols of operands and operations.

3. Formation of new program specification structures that describe the original program with detail down to operations/functions:

- selection from a set of elements of a postfix program specification of a subset of program operation statements/functions;

- continuous numbering of operators;

- continuous numbering of inputs for each operator and outputs;

- input of numeric coding of operator types based on the postfix representation of each operation/function;

- formation for each operator of a generated numerical specification of a set of numbers of its operands and setting its power;

- formation for each operator of a set of its external operators (using the results of the operator’s execution) and setting its power;

- formation, based on the postfix representation of operations/functions, for each operator of the corresponding labels.

4. Formation of new program specification structures that describe the original program with detail down to fragments:

- selection from the set of operators of the main structure of subsets of operators that have the same fragment number value:

- determination for each subset of numbers and types of input and output operators, fixation of their control connections and corresponding control transfer labels;

- formation in numerical format of the main, connected and temporary structure, specifying the types and scheme of control connections of fragments.

The results are the following generated new program specification structures at this level:

- basic structure of fragments: fragment number; fragment labels; types of fragments; pointers to the fragment number; cardinality of the conjugate set for a fragment; pointer to the beginning of the sequence of fragment numbers that form the external set of the fragment; cardinality of the outer set for the fragment; labels of fragments of unconditional transition and conditional transition by value “true”; labels of fragments of a conditional jump based on the value “false” of the program.

- connected structure of fragments: number of lines of the connection structure; a pointer to the continuation of the sequence of fragment numbers forming the conjugate fragment set for the fragment in question; the conjugate set of a fragment for the fragment under consideration; a pointer to the continuation of the sequence of fragment numbers forming the external fragment set for the fragment in question; the outer set of the fragment.

- temporary structure of fragments: number of vertices; graph vertex number; the point in time at which execution of a parallel algorithm instruction begins, interpreted by the vertex corresponding to the temporary parallel circuit.

5. Assessing the correctness of the generated new specification structures of the original program in order to check the equivalence of the text specification of the task program and its representation by new specification structures, i.e. assessment of the equality of the numbers of conjugate and external connections in the main and connected structures of the new specification, assessment of the correspondence of the numbers of operators of various types in the main and connected structures and the original program, assessment of the correspondence of the number of inputs, outputs and their bit sizes in the main and connected structures and in the original program, assessment of the correctness of the semantics of inputs and outputs of operators of the main and connected structures in comparison with the semantics (units of measurement) of the operands and results of executing instructions/functions of the program, assessment of the equivalence of control schemes specified by the main and connected structures and the program text.

6. Expansion of the set of operators and the set of connections between the operators of the main and connected structures of the new specification of the source program through the introduction of additional variables representing the results of intermediate calculations of the right-hand sides of the assignment operators, and the operators corresponding to these variables for writing data into memory, which ensures the transition to constructive main and connected structures of the new specification, displaying the full composition and connections of operators to be executed when solving a problem.

7. Implementation of a cycle across temporary tiers of a new task specification in order to form a set of fragments for the task (beginning of fragmentation).

8. Formation for a set of operators of each tier of a set of operators ranked in descending order of power of their conjugate sets of operators.

9. Formation of the first fragment includes:

- current value of the number of fragments;

- the first fragment includes a set of operators that have the maximum cardinality of the conjugate set;

- the complexity of each fragment is equal to the sum of the powers of the conjugate and external sets of operators included in the fragment.

10. Formation for the current operator belonging to a certain time tier of a set of intersections of conjugate sets and checking the equality of the conjugate set to the empty set in order to verify the fact that the current and output operators use common source data.

11. Checking whether the current number of fragments has reached the specified number and including the current statements in the next fragment.

12. Calculation of the fragmentation complexity for the current fragment by adding the fragmentation complexity of the fragment calculated at the previous stages and the power of the conjugate and external sets of operators included in the fragment.

13. Checking the end of consideration of all operators of temporary tiers.

14. Formation for operators of temporary tiers of sets of intersections of its conjugate set and subsets of operators of each fragment of the current set of generated fragments.

15. Determining the number of the fragment that has the maximum intersection value, including the temporary tier operator in the fragment and calculating the fragmentation complexity (according to step 12) for the current fragment.

16. Estimation of the power of the sets of operators included in each of the fragments in order to ensure uniformity of the number of operators in the fragments and uniformity of resource loading.

17. Determining the number of the fragment that has the minimum cardinality of the operator sets and including the temporary tier operator in the fragment and calculating the fragmentation complexity (according to step 12) for the current fragment.

18. Calculation of the total complexity of task fragmentation, completion of fragmentation.

19. Clustering of fragments in the interests of minimizing the total time of message exchange.

20. Assessment of the current state of the resource of processors and the communication network, which consists in determining the composition of the currently free resource (taking into account the possibility of completing previously started tasks):

- the number and set of numbers of currently free processors that have a busy sign;

- the number and variety of numbers of free communication lines that have a busy sign.

21. Establishing a one-to-one correspondence between clusters of fragments and a subset of currently free processors of computing nodes, assigning a dedicated processor to each cluster of fragments and ensuring minimization of the total length of interprocessor connections.

22. Evaluation of a specific type of processor (superscalar and with a super-long command line).

23. Formation of many fragments for superscalar processors and processors with a very long command line.

24. Assessment of a specific topology (fully connected, ring, hypercube, common bus, lattice-torus, etc.) of the communication network of a computer system.

25. Estimation of the maximum time spent on messaging, i.e. to execute a parallel process with time parameterization in a computing system, which is determined by the amount of CPU time spent and interprocessor message exchange:

- cycle by numbers of fragments, fragmentary main and coherent structures of the new specification;

- determination of the number of the processor implementing the fragment;

- allocation for a fragment of its external fragment set and a set of numbers external to the processor with the number of processors implementing the set of fragments;

- selection from the main and connected structures of the new computer system topology specification of an external set of processor numbers for the main processor with the corresponding number;

- formation of the difference between external sets of processor numbers, if it is greater than or equal to the empty set, then this corresponds to the correct assignment of the main and its external fragments to the necessary topology resource and the time for implementing the exchange is determined as the sum of the total time spent on exchanging messages when solving the problem and the product of time transferring service data to a set of numbers of external processors that implement the external fragment set, when considering all fragments of the task - completing the assessment of time communication costs, in another case, moving on to executing the cycle for the next fragment;

- formation of the difference between external sets of processor numbers, if it is less than an empty set, then this corresponds to the insufficiency for the main processor with the corresponding number of the required number of adjacent processors to execute all fragments of the set and the need to select additional processors from the available resource to implement the remaining unallocated fragments of the set of external fragments for fragments (providing the possibility of a minimal increase in additional time spent on exchange when implementing a fragment):

- a cycle through the numbers of processors forming an external set of processors for the corresponding main processor;

- selection from the main and connected structures of the new specification of the topology of the computer system of an external set of processor numbers for the processor with the corresponding number;

- loop through elements of the external set of the processor with the corresponding number;

- checking the busy condition of the processor with the corresponding number;

- if the processor is busy, transition to the next element of the external set of the processor, in another case - assignment of the current unassigned fragment to the processor with its transfer to the busy state;

- the formation of the current value of the time spent on the implementation of the exchange is defined as the sum of the total time spent on exchanging messages when solving a problem and the product of the time for transferring service data by the sum of the difficulties of data exchange between fragments and their current external fragments;

- formation of parallel fragments with time parameterization of the process of solving a problem by a computer system with determination of the real time costs for parallel execution of the task are determined by the possibilities of combining the execution of various processor operations and message exchange operations in time and take into account: the number of computing nodes and processors; values of durations for performing various types of operations, incl. operations accessing individual processor memory; number of data reception/transmission ports of an arbitrary processor; computer system topology; parallel data processing methods;

- assessment of efficiency indicators of generated fragments with time parameterization of the process of solving a problem by a computer system: sequential and parallel time for solving a problem, time gain, parallelization efficiency indicator, equipment load factor.

26. Development of text specifications of threads of a parallel program with time parametrization includes the development of the following objects:

- text specifications of temporary threads of a parallel program with built-in means (sleep operators) for time synchronization of parallel processes of computing nodes;

- text specifications of temporary threads of a parallel program with built-in tools (send, receive operators), taking into account the time of transmitting/receiving data into the buffers of the sending and receiving parties, the time of receiving sent data into memory;

- structures of the temporary specification of a parallel program in the form of individual texts of program threads of computer system processors with built-in tools (sleep, send, receive operators);

- estimates of the total number of calls to each data required to execute the original program.

Initial data of the stage of development of text specifications of temporary threads of a parallel program: new structures (main, connected, temporary) of a parallel process with temporary parameterization that meets the specified requirements (implementation time of a parallel program on a given resource, processor load factor depending on the topology of the computer system, type and number of processors ); assigning operators to threads or processors; characteristics of the computing system architecture: type of computer system topology, number of computing nodes, number and type of processors, duration values for various types of operations and operations for accessing individual processor memory, number of simultaneous parallel I/O ports, methods of parallel data processing.

The main stages of developing text specifications of temporary threads of parallel program programs include:

- cycle through operator numbers of the main structure;

- determination of the number of the processor implementing the operator;

- formation of a text specification of the operator/operation, including:

1) determination using the main structure of the type of processor instruction;

2) sampling the names of conjugate operators from the coherent structure and recording the text specification of the operator;

3) sampling the names of conjugate operators from the coherent structure, replacing the names of operators with their actual addresses and recording the text specification of the operator;

4) temporary parameterization of operators of processor threads of programs;

5) presentation of text specifications of program threads with time parameterization of each of the processors of the computing node for programs in the form of a set of lines of the following form: numbers of commands of processor program threads; operation class attribute; operation name; addresses of the first and second operands; fragment names; the value of the current discrete time corresponding to the beginning of the implementation of the task operators, containing the specific data used when performing the operation.

27. Assessing the correctness of the results of developing parallel programs with time parameterization, i.e. assessment of the correctness of data types, types of operations/functions on data, connections between data and control operations, correctness of units of measurement of physical quantities, correctness of the start and duration of computational operations/functions and operators of control transfers and synchronization of temporary parallel processes.

28. Compiling the generated parallel code with time parameterization.

Thus, in order to increase the efficiency of digital information processing (reduce the time for solving a problem), temporary program threads should be developed taking into account the requirement to optimize the uniformity of loading of computing nodes in the process of parallel solution of the problem, thereby ensuring the necessary time synchronization of the operation of computing nodes.

New features with significant differences are:

1. Taking into account the architecture of massively parallel computing systems with distributed memory.

2. Taking into account options for task fragmentation.

3. Taking into account the parameter of the start time of execution of fragments/operators of the parallel algorithm.

These signs have significant differences, since they were not found in known methods.

The use of new features, in combination with the known ones, will improve the efficiency of digital information processing by optimizing the uniformity of loading of computing nodes in the process of parallel solution of the problem.

The method of time synchronization of the operation of a massively parallel computing system with distributed memory is implemented as follows.

In fig. Figure 1 shows a diagram of the main components of a massively parallel computing system with distributed memory, consisting of a communication network 3, computing nodes 1, 2, which include input/output interfaces 11, 21, local memories 12, 22, processors with cache memory 13, 23 , network adapters 14, 24, system buses 15, 25, and a host computer 4, which includes an instruction fetch unit 41, a fragmentation unit 42, a control device 43, an arithmetic-logical unit 44, a data memory 45. The computing system contains two computing nodes 1, 2 and host computer 4, configured to optimize task fragmentation and optimize message exchange between computing nodes 1, 2, designed to create parallel code with time parameterization through the communication network 3.

Let's consider the step-by-step implementation of the proposed method of time synchronization of the operation of a massively parallel computing system with distributed memory in the system described above (Fig. 1). The data memory 45 is loaded with serial program code. From data memory 45, control device 43 receives an ordered selection of tokens from the original sequential program and generates token descriptors (step 1). Instruction sampling unit 41 receives an ordered selection of lexeme descriptors and generates a postfix specification for this construction (step 2). Forms new program specification structures with detail down to operations/functions (step 3). Forms new program specification structures with detail down to fragments (step 4). The control device 43 checks the equivalence of the textual specification of the task program and its representation by the new specification structures (step 5). Expands the sets of operators and sets of connected structures of the new specification of the original program (step 6). The fragmentation block 42 implements a loop through the temporary tiers of the new task specification (step 7). For the set of operators of each tier, it generates sets of operators ranked in descending order of power of their conjugate sets of operators (step 8). Forms the first fragment (step 9). Forms for the current operator a set of intersections of conjugate sets belonging to a certain time tier and checks the equality of the conjugate set to the empty set (step 10). Checks whether the current number of fragments has reached the specified number and includes the current statements in the next fragment (step 11). Arithmetic logic unit 44 calculates the fragmentation complexity for the current fragment (step 12). The control device 43 checks that all temporary tiers have been processed (step 13). For operators of temporary tiers, it generates sets of intersections of its conjugate set and subsets of operators of each fragment of the current set of generated fragments (step 14). The fragmentation unit 42 determines the number of the fragment having the maximum intersection value, includes the temporary tier operator in the fragment, and the arithmetic logic unit 44 calculates the fragmentation complexity for the current fragment (step 15). The control device 43 evaluates the power of the sets of statements included in each of the fragments (step 16). The fragmentation block 42 determines the number of the fragment that has the minimum value of the cardinalities of the operator sets, includes the temporary tier operator in the fragment, and the arithmetic-logical unit 44 calculates the fragmentation complexity for the current fragment (step 17). Calculates the overall difficulty of task fragmentation (step 18). The control device 43 clusters the fragments (step 19). Evaluates the current state of the resource of processors with cache memory 13 and communication network 3 (step 20). Establishes a one-to-one correspondence between clusters of fragments and a subset of currently free processors with cache memory 13, 23 of computing nodes 1, 2 (step 21). The control device 6 evaluates the specific type of cache processors 13, 23 (step 22). Generates multiple fragments for superscalar and super-long-line processors (step 23). Evaluates the specific topology type of communication network 3 (step 24). Estimates the maximum time spent exchanging messages (step 25). Instruction sampling unit 41 develops text specifications of parallel program threads with time parametrization, taking into account the time of data transmission/reception into the buffers of the sending and receiving sides and adds built-in means (sleep, send, receive operators) of time synchronization to the code for each processor with cache memory 13 , 23 computing nodes 1, 2 (step 26). The control device 43 evaluates the correctness of the development results of parallel programs with time parameterization (step 27) and, through the communication network 3, sends a command to processors with cache memory 13, 23 to compile the created parallel code with time parameterization (step 28). Below is an example of the generated parallel code with time parameterization (Fig. 2).

Thus, the proposed method will reduce the time required to solve a problem on two computing nodes by up to 24% when operating a massively parallel computing system with distributed memory, that is, increasing the efficiency of digital information processing due to time synchronization of the operation of a massively parallel computing system with distributed memory.

Information sources

1. Yakovlev S.V., Safonov I.V., Bykova T.V. Method of constructing a program. Patent for invention No. 2406112, Bulletin. No. 34, 2010 (analogue).

2. Viktorov D.S., Brezhnev D.Yu., Tolmachev A.A., Kalachnikov A.S., Yakunina G.R. A method for automatically creating a parallel program with time parameterization of multiprocessor computer systems with identical memory access. Patent for invention No. 2786347, Bulletin. No. 35, 2022 (prototype).

Claims (1)

A method for time synchronization of the operation of a massively parallel computing system with distributed memory, which consists in the fact that to implement the method, a communication network, computing nodes, which include input/output interfaces, local memories, processors with cache memory, network adapters, system buses and a host - a computer, which includes an instruction fetch unit, a fragmentation unit, a control device, an arithmetic-logical unit, and data memory, performs the following operations: an ordered selection of tokens of the original sequential program is obtained and their descriptors are formed for the tokens; obtaining an ordered selection of lexeme descriptors from the set of descriptors generated at the previous stage; form new program specification structures with detail down to operations/functions; form new program specification structures with detail down to fragments; check the equivalence of the text specification of the task program and its representation by new specification structures; expand the sets of operators and sets of connections between operators of the main and connected structures of the new specification of the original program; implement a cycle along the temporary tiers of the new task specification; for the set of operators of each tier, sets of operators are formed, ranked in descending order of power of their conjugate sets of operators; form the first fragment; for the current operator belonging to a certain time tier, sets of intersections of conjugate sets are formed and the equality of the conjugate set to the empty set is checked; check whether the current number of fragments has reached the specified number and include the current operators in the next fragment; calculate the fragmentation complexity for the current fragment; check the end of consideration of all operators of temporary tiers; form for the operators of temporary tiers a set of intersections of its conjugate set and a subset of operators of each of the fragments of the current set of generated fragments; determining the number of the fragment having the maximum intersection value, including the temporary tier operator in the fragment, and calculating the complexity of fragmentation for the current fragment; estimating the power of the sets of operators included in each of the fragments; determine the number of the fragment that has the minimum value of the cardinality of the operator sets, include the temporary tier operator in the fragment and calculate the complexity of fragmentation for the current fragment; calculate the overall complexity of task fragmentation; cluster fragments; evaluate the current state of the resource of processors with cache memory and the communication network; establish a one-to-one correspondence between clusters of fragments and a subset of currently free processors with cache memory of computing nodes; evaluate a specific type of processor with cache memory; form multiple fragments for superscalar processors and processors with super-long command line; evaluate the specific type of communication network topology; estimate the maximum time spent on messaging; develop text specifications of parallel program threads with time parametrization, taking into account the time of data transmission/reception into the buffers of the transmitting and receiving sides and adding built-in time synchronization tools to the code for each processor with cache memory of the computing nodes; evaluate the correctness of the results of developing parallel programs with time parameterization; compile the generated parallel code with time parameterization.

Images