RU124009U1 - PARALLEL FLOW COMPUTING SYSTEM - Google Patents

Abstract

1. A parallel streaming computing system comprising a token switch and a plurality of actuators, a plurality of vector actuators and a plurality of associative memory modules, each of the actuators being directly connected to its corresponding associative memory module, forming one computational core, and each from vector actuators is directly connected to its corresponding associative memory module, forming one vector computation tion yadro.2. The computing system according to claim 1, characterized in that the token switch is configured to receive tokens, or multi-input tokens, or vector tokens, transmit each token, or multi-input token, or vector token to all computing cores and vector computing cores at once, or in accordance with with the number of the computing core or vector computing core indicated in one of the fields of the token, or multi-input token, or vector token, transfer the token, or multi-input token, or vector token to the ac module otsiativnoy computational kernel memory or computational kernel vector corresponding to said number of computational kernel or vector computing yadra.3. The computing system according to claim 1, characterized in that each associative memory module is configured to receive tokens, or multi-input tokens, or vector tokens from the switch, to compare the token, or multi-input token, or vector token, taking into account masking with previously received tokens, or multi-input tokens, or

Description

Application area

The claimed utility model relates to computing, in particular to computing systems based on a computational model with data flow control.

State of the art

Known, selected as the closest analogue, is an information processing system using an approach based on a data flow control containing one or more associative memory modules, each of which is capable of storing tokens entering them containing a data field and a key field, to search among stored tokens tokens in which the key matches the key of the incoming token, remember the incoming token in the associative memory modules in case of an unsuccessful search, generate a pair of data from the post a token entering the associative memory and a token found in the associative memory upon successful search and sending the generated data pair to any of the actuators, one or more actuators configured to execute instructions stored in the instruction memory accessible to the actuator at the addresses specified in pair of data, generate a token containing the data field and the key field, in which the result of the instruction is placed in the data field, and information about the instruction in the key field and, for which this result is a parameter, calculate a hash function that determines the number of the associative memory module, send the generated token to the associative memory module with any number in accordance with the calculated hash function of the token switch to direct tokens from any executive device to any module associative memory, and a switch of data pairs for sending bunks of data from any associative memory module to any actuator, with each associative memory module with the ability to store tokens in any of its K + 1 sections, each of which has its own fixed size and numbered from 0 to K, receive a token for storage from executive devices, which in this case is called the input, calculate the whole from the information in the input token C, which is in the range from 0 to K-1, to perform the “token search” by key in the input token simultaneously in sections with numbers C and K by comparing the key of the input token with the keys of all stored in the token section, if the operation “by ck of the token "to send the found token to the device performing the operation" generating a data pair "and sending the generated data pair to the executive device through the switch of the data pairs, if the operation" token search "is unsuccessful, write the token to the section with the number C, and if if the number of tokens in this section is equal to the maximum possible for this section, record the token in the section with number K (publication RU 2360279 C2, cl. IPC G06F 7/76, publ. June 27, 2006 Bull. No. 18).

The presence of a common switch of data pairs makes the hardware implementation of a scalable computing system impossible; in addition, there is a limitation on the amount of input data of an executable instruction (program) of an executive device equal to two. The used token formats and data pairs do not allow the formation of multi-input structures. There is no possibility of implementing vector operations, which does not allow the transfer of data arrays and, therefore, reduces the speed of the system as a whole.

Thus, the disadvantages of the known system are its low performance, the inability to scale the computing system, as well as the limitation on the number of input data of the node program, the absence of vector operations, which does not allow to achieve high system performance and its speed.

Utility Model Disclosure

The technical result that can be obtained in the claimed utility model is to eliminate the above disadvantages.

The technical result is achieved in that the parallel streaming computing system comprises a token switch and a plurality of actuators, a plurality of vector actuators and a plurality of associative memory modules, each of the actuators being directly connected to its corresponding associative memory module, forming one computational the core, and each of the vector actuators is directly connected to its corresponding module associatively th memory, forming one vector computing core.

And also the fact that the token switch is configured to receive tokens, or multi-input tokens, or vector tokens, transfer each token, or multi-input token, or vector token to all computing cores and vector computing cores at once, or, in accordance with one from the fields of the token, or multi-input token, or vector token with the number of the computing core or vector computing core, transfer the token, or the multi-input token, or vector token to the associative memory module of the computing nucleus or vector computational kernel corresponding to said computational kernel number or vector computing nucleus.

And also the fact that each associative memory module is capable of accepting tokens, or multi-input tokens, or vector tokens from a token switch, to compare a token, or a multi-input token, or a vector token, taking into account masking with previously received tokens, or multi-input tokens, or with vector tokens and, if coincidentally, perform actions on the token, or multi-input token, or vector token in accordance with the operation code of matching tokens, or multi-input tokens, or vector token c, generate and transmit to the actuator a data pair or multi-input packet containing respectively the data of tokens, or multi-input tokens, or vector tokens, or generate and transmit to the vector executive device a vector packet containing respectively data of tokens, or multi-input tokens, or vector tokens , and store the received tokens, or multi-input tokens, or vector tokens.

And also the fact that each actuator is configured to receive a pair of data or a multi-input packet, extract information from them about the address of the program stored in the actuator and which should be performed on the data of a pair of data or a multi-input packet, execute the specified program on the data of the data pair or multi-input packet, generate new tokens, or multi-input tokens, or vector tokens and transfer them to the token switch.

And also the fact that each vector execution unit is configured to receive a vector package, extract information from it about the address of the program stored in the vector execution unit and which should be performed on the vector package data, execute the specified program on the vector package data, generate new tokens, or multi-input tokens, or vector tokens and transfer them to the token switch.

Description of drawings

The claimed parallel streaming computing system is illustrated using the schemes presented in figures 1-6.

Figure 1 presents the structural diagram of the claimed parallel streaming computing system.

Figure 2 shows the structure of the field "token signs" of the token, multi-input token, vector token.

Figure 3 shows the structure of the field "package signs" multi-input package, vector package.

Figure 4 shows a diagram of the formation of a vector token, which is used in the claimed parallel streaming computing system.

Figure 5 shows a diagram of the formation of a multi-input package, which is used in the claimed parallel streaming computing system.

Figure 6 shows a diagram of the formation of a vector package, which is used in the claimed parallel streaming computing system.

Utility Model Implementation

The claimed parallel streaming computing system shown in figure 1, contains many modules 5 of associative memory, many actuators 3, many vector actuators 10 and switch 6 tokens. Each of the actuators 3 is directly connected to its associated module 5 of associative memory, forming one computing core 1, and each of the vector actuators 10 is directly connected to its corresponding module 5 of associative memory, forming a vector computing kernel 9. Moreover, many modules 5 are associative memory, many actuators 3 and many vector actuators 10 are connected to the switch 6 tokens.

The lines 2a connecting the token switch 6 with the associative memory modules 5, and the lines 2b connecting the actuators 3 and the vector actuators 10 with the token switch 6 transmit tokens, or multi-input tokens, or vector tokens. On lines 4a connecting the associative memory modules 5 with actuators 3, data pairs or multi-input packets are transmitted. The lines 4b connecting the modules 5 of the associative memory with vector actuators 10 transmit vector packets.

Both tokens, multi-input tokens, and vector tokens represent a data structure consisting of the “token attributes” field (see FIG. 2), the “data” field and the “key” field, while the “key” field includes the “context” field ”And the“ node address ”field. In the“ token attributes ”field shown in FIG. 2, the signs of global transmission are written to the associative memory modules 3 (MAP Glo) or the number of the associative memory module 3 (MAP number) to which corresponding token, token priority, mask number, operation code (Kop), and for multi-input current and a vector token is also the number of inputs (KBX) and the input number (NVH). The priority of the token indicates the order in which the token is processed if there is more than one token in any of the modules 3 of the associative memory, as well as in the switch 6. If there is more than one token in any of the modules 3 of associative memory, as well as in switch 6, the token with a higher priority is first processed. The operation code is an internal command of associative memory module 3, which must be executed on this token in case of coincidence of two tokens, taking into account masking. In the “data” field, data is recorded for subsequent processing, in the “key” field, information is recorded that indicates what actions should be performed in the future on the result, while information in the context field is written about the context in which these actions be produced, and in the field "node address" write the address of the program that must be executed on the data.

In this description, “multi-input tokens” means tokens by which a multi-input packet is formed. Unlike tokens, multi-input tokens contain two additional fields located in the “token attributes” field - “input number” (Нвх) and “number of inputs” (Квх). The “entry number” indicates which of the fields of the “given” resulting packet the given of this multi-input token should fall into, and the “number of inputs” serves to control and indicates how much data should be in the multi-input packet.

Under the "vector token", presented in figure 4, in this description refers to the data structure generated on the basis of multi-input tokens and containing the following fields: field "token attributes", field "key", many fields "this". In this case, the “key” field includes the “context” field and the “node address” field (Auzl.).

The “context” field is formed, for example, by the logical addition of all bits of the zero “context” of all the multi-input tokens participating in the formation of the specified vector token, taking into account the mask of each of them. If the “context” field is not masked for matching multi-input tokens, then this means that it is the same for all multi-input tokens, and, accordingly, for a vector token, it will be the same.

The “node address” field of a vector token is formed, for example, by the logical addition of all bits of the “node address” field of all multi-input tokens participating in the formation of a vector token taking into account the mask of each of them. If the “node address” field is not masked for matching multi-input tokens, then this means that it is the same for all multi-input tokens, and, accordingly, for a vector token, it will be the same.

The “data” fields of the vector token are filled with the data of all the multi-input tokens participating in the formation of the vector token, depending on the value of the “input number” field of the “token attributes” field of the multi-input token.

It should be noted that the data of the “data” fields of the vector token are a one-dimensional array of the same type of data placed regularly.

Under the "multi-input packet", presented in figure 5, in this description refers to a data structure containing the following fields: field "package features" field "key" field "this", which should be at least one. In this case, the "key" field includes the "context" field and the "node address" field,

In the field "signs of the package", presented in figure 3, is recorded a sign of priority and a number that determines the amount of data in the package.

The “context” field is formed, for example, by the logical addition of all bits of the “context” field of all tokens, or multi-input tokens, or vector tokens participating in the formation of the package, taking into account the mask of each of them. If the “context” field is not masked for matching tokens, then this means that it is the same for all tokens, and, accordingly, the package will have the same.

The "node address" field is formed, for example, by the logical addition of all the bits of the "node address" field of all tokens participating in the formation of the packet, taking into account the mask of each of them. If the “host address” field is not masked for matching tokens, then this means that it is the same for all tokens, and, accordingly, for the packet it will be the same.

The “data” fields of the packet are filled with the data of all tokens participating in the formation of the packet, depending on the token number in the multi-input packet.

Under the "vector package", presented in Fig.6, in this description refers to a data structure generated on the basis of vector tokens, or multi-input tokens, or tokens, and containing the following fields: field "package signs" field "key", many fields "Given." In this case, the “key” field includes the “context” field and the “node address” field (Auzl.).

In the field “signs of the package” is recorded a sign of priority and a number that determines the amount of data in this vector package.

The “context” field is formed, for example, by the logical addition of all the bits of the “context” field of all vector tokens participating in the formation of the specified vector packet, taking into account the mask of each of them. If the “context” field is not masked for coincident vector tokens, then this means that it is the same for all vector tokens, and, accordingly, for a vector package it will be the same.

The “node address” field of a vector packet is formed, for example, by the logical addition of all the bits of the “node address” field of all vector tokens participating in the formation of a vector packet taking into account the mask of each of them. If the “node address” field is not masked for matching vector tokens, then this means that it is the same for all vector tokens, and, accordingly, for a vector package it will be the same.

The “data” fields of the vector packet are filled with the data of all vector tokens participating in the formation of the vector packet, depending on the value of the “entry number” field of the “token attributes” field in the vector token.

It should be noted that the data of the “data” fields of the vector packet are a one-dimensional or multidimensional array of the same type of data placed regularly.

The claimed parallel streaming computing system operates as follows.

At the beginning of the parallel streaming computing system, the operator sets the data transfer mode in the form of tokens, or multi-input tokens, or vector tokens from the host machine 8 to the parallel streaming computing system, and receiving data in the form of tokens, or multi-input tokens, or vector tokens 8 from a parallel streaming computing system. Data input / output operations from the host machine 8 are carried out by means of the input / output unit 7.

Tokens, or multi-input tokens, or vector tokens, are received from the host 8 through the input / output unit 7 to the token switch 6, which accepts tokens, or multi-input tokens, or vector tokens. Then, if there is a global token in the “token signs” field of the token, or multi-input token, or vector token, then the token switch 6 transmits the specified received token, or multi-input token, or vector token along the token transmission lines 2a to all computing cores 1 and to vector computing cores 9. In the event that in the field “token attributes” of a token, or a multi-input token, or a vector token, the number of computing core 1 or the number of vector computing core 9 is indicated, to which the this token, or a multi-input token, or a vector token, the token switch 6 transmits the specified received token, or a multi-input token, or a vector token via the token transfer line 2a to the associative memory module 5 of the computing core 1 or the vector computing core 9 corresponding to that indicated in the field “Token signs” of a token, or a multi-input token or vector token, to the number of computing core 1 or vector computing core 9.

One or all associative memory modules 5 receive the indicated token, or a multi-input token, or a vector token from the switch 6 of tokens. The first of these tokens, without an object for comparison, is written to associative memory module 5 in anticipation of receiving other tokens. Then one or all associative memory modules 5 receive the next token, or multi-input token, or vector token from token switch 6, and compare the received specified token, or multi-input token, or vector token, taking into account masking with previously received tokens, or multi-input tokens, or vector tokens. In this case, the “key” field or the “node address” and “context” fields are compared. If they coincide with the masking of the “key” fields or the “node address” and “context” fields, they perform the actions in accordance with the operation code of the matching tokens on the received tokens, or multi-input tokens, or vector tokens. In this case, a data pair or multi-input packet is formed in computing cores 1, or a vector packet is formed in vector computing cores 9, or the indicated received token, or multi-input token, or vector token is deleted, or a stored token, or multi-input token, or vector token is deleted , or save the specified received token, or multi-input token, or vector token.

If a token, or a multi-input token, or a vector token is directed to the computing core 1, the associative memory module 5 generates and transmits to the actuator 3 via a transmission line 4a a data pair or a multi-input packet containing data of at least one token or multi-input token , or vector token.

If the token, or multi-input token, or vector token is directed to the vector computing core 9, the associative memory module 5 generates and transmits to the vector executive device 10 via the transmission line 4b a vector packet containing data of at least one token or multi-input token, or vector token.

When forming a multi-input package or vector package, if tokens or multi-input tokens or vector tokens coincide, the “package signs” field is filled on the basis of the “token attributes” fields of the corresponding tokens, or multi-input tokens, or vector tokens, the “node address” field and “Context” are formed from the corresponding fields of tokens, or multi-input tokens, or vector tokens by their mutual logical addition taking into account masking. The "data" fields contain data from the "given" fields of the corresponding tokens, or multi-input tokens, or vector tokens, in accordance with the "input number" field of the token, or multi-input token, or vector token. Figure 5 presents a diagram of the formation of a multi-input packet, which is formed in the module 5 of associative memory from a group of, for example, multi-input tokens. Figure 6 presents a diagram of the formation of a vector package, which is generated in the module 5 of associative memory from a group of, for example, vector tokens.

The actuator 3 receives a data pair or multi-input packet on line 4a from the associative memory module 5, and the vector actuator 10 receives a vector packet on line 4b from the associative memory module 5.

The actuator 3 extracts from the data pair or from the multi-input packet information about the address of the program, which is stored in the executive device 3 and which should be performed on the data of the data pair or multi-input packet. Then, the actuator 3 starts and executes the specified program on the specified data. As a result of the execution of this program, new data is obtained, based on which the actuator 3 generates new tokens, or multi-input tokens, or vector tokens, and transfers the indicated tokens, or multi-input tokens, or vector tokens to the token switch 6 on line 2b.

Vector execution device 10 extracts from the vector package information about the address of the program, which is stored in the executive device 10 and which should be performed on the data of the vector package. Then the vector actuator 10 starts and executes the specified program on the specified data. As a result of the execution of this program, new data is obtained, on the basis of which the vector executive device 10 generates new tokens, or multi-input tokens, or vector tokens, and transfers the indicated tokens, or multi-input tokens, or vector tokens to the token switch 6 on line 2b.

Further, the token switch 6 transmits the received tokens, or multi-input tokens, or vector tokens via lines 2a, or immediately to all computing cores 1 and vector computing cores 9, if there is a global token in one of the fields of the token, or multi-input token, or vector token either to computing core 1 or vector computing core 9, if in one of the fields of the token, or multi-input token, or vector token, the number of computing core 1 or vector computing core 9 is indicated, in which s transmitting said token or multiinput token or token vector.

The data processing process ends after the actuators 3 and vector actuators 10 transmit the final result of the data processing to the token switch 6, which through the input / output unit 7 will issue them to the host machine 8.

Thus, due to the fact that each of the actuators and vector actuators are directly connected to the corresponding associative memory module, forming one computational core and vector computational core, respectively, the performance of the computing system is increased, the possibility of scaling the computing system appears, and the restriction is also removed by the amount of input data of the node program, and also significantly increases the system performance and its speed .

Claims (5)

1. A parallel streaming computing system comprising a token switch and a plurality of actuators, a plurality of vector actuators and a plurality of associative memory modules, each of the actuators being directly connected to its corresponding associative memory module, forming one computational core, and each from vector actuators is directly connected to its corresponding associative memory module, forming one vector computation tion core. 2. The computing system according to claim 1, characterized in that the token switch is configured to receive tokens, or multi-input tokens, or vector tokens, transmit each token, or multi-input token, or vector token to all computing cores and vector computing cores at once in accordance with the number of the computing core or vector computing core indicated in one of the fields of the token, or multi-input token, or vector token, transfer the token, or multi-input token, or vector token to module a sociative memory of the computing core or vector computing core corresponding to the specified number of the computing core or vector computing core. 3. The computing system according to claim 1, characterized in that each associative memory module is configured to receive tokens, or multi-input tokens, or vector tokens from the switch, to compare a token, or multi-input token, or vector token, taking into account masking with previously accepted tokens, or multi-input tokens, or vector tokens, and if matched, perform actions on the token, or multi-input token, or vector token in accordance with the operation code of the matching tokens, or multi-input tokens, or vector tokens, generate and transmit to the executive device a data pair or multi-input packet containing respectively the data of tokens, or multi-input tokens, or vector tokens, or generate and transfer to the vector executive device a vector packet containing respectively data of tokens, or multi-input tokens , or vector tokens, and store the received tokens, or multi-input tokens, or vector tokens. 4. The computing system according to claim 1, characterized in that each actuator is configured to receive a pair of data or a multi-input packet, extract information from them about the address of the program stored in the actuator, and which should be performed on the data of a pair of data or a multi-input packet , execute the specified program on the data of a data pair or multi-input packet, generate new tokens, or multi-input tokens, or vector tokens and transfer them to the token switch. 5. The computing system according to claim 1, characterized in that each vector execution unit is configured to receive a vector package, extract information from it about the address of the program stored in the vector execution unit, and which should be performed on the data of the vector package, execute the specified program over the data of the vector packet, generate new tokens, or multi-input tokens, or vector tokens and transfer them to the token switch.

Images