KR20210106929A - Apparatus for distributed processing through remote direct memory access and method for the same - Google Patents

Abstract

According to one embodiment of the present invention, provided is a distributed processing apparatus through remote direct memory access, comprising: a communication unit that directly accesses remote memories provided in a plurality of shared memory servers constituting a shared memory server cluster to transmit and receive data; a shared memory access management unit that allocates a local shared memory area to a memory by the same size as a shared memory buffer and synchronizes the local shared memory area with the shared memory buffer for the shared memory buffer configured with shared memory buffer segments allocated from the remote memories; a memory mapping table management unit that manages a memory mapping table between the shared memory buffer and the local shared memory area; and a calculation unit that performs a calculation given for the local shared memory area.

Description

Distributed processing device and method through remote direct memory access

The present invention relates to a distributed processing apparatus and method through remote direct memory access, and more particularly, to a distributed processing apparatus and method for directly accessing a virtual shared memory buffer composed of memories of shared memory servers in a distributed processing framework it's about

Distributed parallel processing is to perform large-scale data analysis quickly by using multiple computational resources in parallel at the same time. Processes that are distributed and executed in parallel on multiple computation nodes must share data with each other, and MPI (Message Passing Interface) is a representative data sharing method. However, in the case of continuously updating and referencing the entire distributed processing data asynchronously between all distributed processing processes rather than passing only a part of the distributed processing data in the form of message passing between some processes, it is better to share it in the form of shared memory rather than the MPI method. It is advantageous.

Processes performing distributed processing on a distributed processing platform must frequently transmit and receive large-scale shared data with each other, and the resulting communication overhead occupies a very high proportion in the overall distributed processing performance or processing time. The higher the communication overhead, the longer the waiting time of the computational processors (eg, CPU, GPU, etc.) of the computation node, and this appears as a decrease in resource utilization. The reason for the high communication overhead is that most communication protocol stacks, including TCP/IP, process the messages sent by the application process through multi-level protocol layers, and the memory copy occurs more than once during protocol processing. Therefore, it is required to reduce communication overhead due to distributed processing through remote direct memory access (RDMA).

The above-mentioned background art is technical information possessed by the inventor for the derivation of the present invention or acquired in the process of derivation of the present invention, and cannot necessarily be said to be a known technique disclosed to the general public prior to the filing of the present invention.

Domestic Patent Publication No. 10-2006-0009244

It is an object of the present invention to provide a distributed processing apparatus and method through remote direct memory access that share distributed processing data through remote direct memory access.

In addition, an object of the present invention is to configure a shared memory buffer by clustering a plurality of shared memory servers and allocating shared memory buffer segments from each of the shared memory servers, and a distributed processing apparatus for remote direct memory access to the shared memory buffer and its to provide a way

An embodiment of the present invention, a communication unit for transmitting and receiving data by directly accessing remote memories provided in a plurality of shared memory servers constituting the shared memory server cluster; For a shared memory buffer composed of shared memory buffer segments allocated from the remote memories, a local shared memory area of the same size as the shared memory buffer is allocated in memory, and the shared memory buffer and the local shared memory area are synchronized a shared memory access management unit; a memory mapping table management unit for managing a memory mapping table between the shared memory buffer and the local shared memory area; And it provides a distributed processing device through remote direct memory access, characterized in that it comprises a calculation unit for performing a given operation on the local shared memory area.

In this case, the shared memory access management unit directly creates the shared memory buffer to share shared memory buffer information with other distributed processing devices constituting the same distributed processing framework, or the shared memory created by the other distributed processing device Receive shared memory buffer information corresponding to the buffer.

In this case, the shared memory access management unit calculates the size of the shared memory buffer segments corresponding to each of the shared memory servers, and requests the shared memory servers to generate and allocate the shared memory buffer segments to the shared memory. You can create a buffer.

At this time, when the data of the local shared memory area is changed by the operation, the shared memory access management unit copies the data of the local shared memory area to the shared memory buffer to synchronize data with the remote memories, and When the data of the shared memory buffer is changed by the distributed processing devices, the data of the shared memory buffer may be copied to the local shared memory area to synchronize data with the remote memories.

In this case, the shared memory access management unit may perform a data accumulation operation between two or more shared memory buffers.

In this case, the shared memory access management unit requests the shared memory servers for an accumulation operation, receives the result of performing the accumulation operation on the shared memory buffer segments from the shared memory servers, and performs the data accumulation operation. can

In this case, the shared memory access management unit requests the shared memory servers to release and delete the shared memory buffer segments, and releases the local shared memory area upon receiving the results of the request to release and delete the shared memory buffer segments. and deletion to end use of the shared memory buffer, and the memory mapping table manager may delete the memory mapping table when the use of the shared memory buffer is terminated.

Another embodiment of the present invention provides a communication unit for transmitting and receiving data to and from a plurality of distributed processing devices through remote direct memory access; a memory directly accessible to the distributed processing devices; It provides a shared memory server, characterized in that it comprises a shared memory management unit constituting a shared memory buffer with other shared memory servers constituting the same shared memory server cluster.

In this case, the shared memory buffer may be synchronized using a memory mapping table and a local shared memory area allocated to each of the distributed processing units in the same size as the shared memory buffer.

At this time, the shared memory management unit receives a request for creation and allocation of the shared memory buffer segment together with size information of the shared memory buffer segment for creating the shared memory buffer from the distributed processing device, and sets the shared memory buffer segment You can configure the shared memory buffer by creating and allocating it.

In this case, when data of a specific local shared memory area is changed by operation, the shared memory buffer may be synchronized with the changed data of the local shared memory area, and may be synchronized with the other local shared memory areas with the changed data.

In this case, the shared memory management unit receives a data accumulation operation request between two or more shared memory buffers from the distributed processing device, and performs an accumulation operation on the shared memory buffer segments that are the target of the data accumulation operation, A result may be returned to the distributed processing unit.

At this time, the shared memory management unit receives the request for releasing and deleting the shared memory buffer segment sent by the distributed processing device to end the use of the shared memory buffer, releasing and deleting the shared memory buffer segment, and displaying the result By returning to the distributed processing unit, the distributed processing unit may release and delete the local shared memory area and delete the memory mapping table.

Another embodiment of the present invention, with respect to a shared memory buffer consisting of shared memory buffer segments allocated from remote memories provided in a plurality of shared memory servers constituting a shared memory server cluster, the shared memory buffer in memory allocating a local shared memory area of the same size as ; managing a memory mapping table between the shared memory buffer and the local shared memory area; synchronizing the shared memory buffer and the local shared memory region by directly accessing the remote memories to transmit and receive data; and performing a given operation on the local shared memory region, providing a distributed processing method through remote direct memory access.

In this case, the shared memory buffer is directly created to share shared memory buffer information with other distributed processing devices constituting the same distributed processing framework, or a shared memory corresponding to the shared memory buffer created by the other distributed processing device. The method may further include receiving buffer information to obtain the shared memory buffer information.

In this case, the step of obtaining the shared memory buffer information comprises: calculating the size of the shared memory buffer segments corresponding to each of the shared memory servers; and requesting the shared memory servers to generate and allocate the shared memory buffer segments to create the shared memory buffer.

In this case, in the synchronizing step, when the data of the local shared memory area is changed by the operation, the data of the local shared memory area is copied to the shared memory buffer to synchronize data with the remote memories, and the other distributed When the data of the shared memory buffer is changed by the processing devices, the data of the shared memory buffer may be copied to the local shared memory area to synchronize data with the remote memories.

In this case, the method may further include performing a data accumulation operation between two or more shared memory buffers.

In this case, performing the data accumulation operation may include: requesting the accumulation operation to the shared memory servers; and receiving a result of performing an accumulation operation on the shared memory buffer segments from the shared memory servers.

In this case, the shared memory servers request release and deletion of the shared memory buffer segments, and release and delete the local shared memory area upon receiving the results of the request for release and deletion of the shared memory buffer segments, and the memory The method may further include terminating use of the shared memory buffer by deleting the mapping table.

According to the present invention, communication overload occurring during distributed processing can be effectively reduced by sharing distributed processing data through remote direct memory access by using the distributed processing apparatus and method through remote direct memory access.

In addition, the present invention configures a shared memory buffer and distributes processing by clustering a plurality of shared memory servers and allocating shared memory buffer segments from each shared memory server by a distributed processing apparatus and method through remote direct memory access. By remote direct memory access to the shared memory buffer by the device, memory data can be efficiently managed without a separate synchronization operation between the shared memory servers.

1 is a diagram showing the configuration of a distributed processing system through remote direct memory access according to an embodiment of the present invention.
FIG. 2 is a block diagram illustrating an example of a distributed processing apparatus through remote direct memory access illustrated in FIG. 1 .
3 is a block diagram illustrating an example of the shared memory server shown in FIG. 1 .
4 is a diagram illustrating a method of configuring a shared memory buffer according to an embodiment of the present invention.
5 is a flowchart illustrating a distributed processing method through remote direct memory access according to an embodiment of the present invention.
6 is an operation flowchart illustrating an example of a step of creating and allocating the shared memory buffer shown in FIG. 5 .
7 is an operation flowchart illustrating an example of a step of releasing and deleting the shared memory buffer shown in FIG. 5 .
8 is an operation illustrating a data accumulation operation method of shared memory buffers according to an embodiment of the present invention.
9 is an operation flowchart illustrating a data accumulation operation method of the shared memory buffers shown in FIG. 8 .

Since the present invention can apply various transformations and can have various embodiments, specific embodiments are illustrated in the drawings and described in detail. Effects and features of the present invention, and a method of achieving them, will become apparent with reference to the embodiments described below in detail in conjunction with the drawings. Here, repeated descriptions, well-known functions that may unnecessarily obscure the gist of the present invention, and detailed descriptions of configurations will be omitted. The embodiments of the present invention are provided in order to more completely explain the present invention to those of ordinary skill in the art. Accordingly, the shapes and sizes of elements in the drawings may be exaggerated for clearer description.

However, the present invention is not limited to the embodiments disclosed below, but all or some of the embodiments may be selectively combined and implemented in various forms. In the following embodiments, terms such as first, second, etc. are used for the purpose of distinguishing one component from another, not in a limiting sense. Also, the singular expression includes the plural expression unless the context clearly dictates otherwise. In addition, terms such as include or have means that the features or components described in the specification are present, and do not preclude the possibility that one or more other features or components will be added.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings, and when described with reference to the drawings, the same or corresponding components are given the same reference numerals, and the overlapping description thereof will be omitted. .

1 is a diagram illustrating the configuration of a distributed processing system 100 through remote direct memory access according to an embodiment of the present invention.

Referring to FIG. 1 , in the distributed processing system 100 through remote direct memory access according to an embodiment of the present invention, the distributed processing devices 110 through a plurality of remote direct memory accesses are remote direct memory access (RDMA: Remote Direct Memory Access) is interconnected with a plurality of shared memory servers 120 through the support network (130). Here, the shared memory servers 120 constitute one shared memory server cluster 140 .

Distributed processing apparatus 110 through remote direct memory access according to an embodiment of the present invention consists of shared memory buffer segments allocated from remote memories provided in a plurality of shared memory servers constituting a shared memory server cluster. For the shared memory buffer, allocating a local shared memory area equal to the size of the shared memory buffer in memory, and synchronizing the shared memory buffer and the local shared memory area using a memory mapping table between the shared memory buffer and the local shared memory area. characterized in that Then, a given or input operation is processed for the local shared memory area.

Here, the distributed processing devices 110 through a plurality of remote direct memory accesses may be included in one distributed processing framework. In addition, the distributed processing devices 110 through one or more remote direct memory access may be configured as one computation node to provide a computation function.

At this time, the distributed processing unit 110 through remote direct memory access is controlled by the master distributed processing unit that handles initialization and control of distributed processing tasks in the distributed processing framework, or the master distributed processing unit, and the slave responsible for calculation It can be divided into a distributed processing unit or a worker distributed processing unit.

At this time, the master distributed processing unit creates a shared memory buffer for storing data in the shared memory servers 120 and transmits the shared memory buffer information to the slave distributed processing units so that all the distributed processing units 110 share the shared memory. It allows access to the same memory segment area on the server 120 . Here, the shared memory buffer information may include the total size of the shared memory buffer, a shared memory buffer generation key, shared memory buffer segment information generated for each shared memory server, and the like.

In this case, the distributed processing device 110 through remote direct memory access copies the data of the local shared memory area to the shared memory buffer when the data of the local shared memory area is changed by operation to synchronize the data with the remote memories. have. In addition, when the data of the shared memory buffer is changed by the operation of the distributed processing unit 110 through another remote direct memory access, the data of the shared memory buffer is copied to the local shared memory area to synchronize the data with the remote memories. have.

The shared memory server 120 is a device that provides shared memory in a distributed processing framework.

Here, the plurality of shared memory servers 120 may constitute one shared memory server cluster. In addition, one or more shared memory servers 110 may be configured as one memory service node to provide a shared memory service.

At this time, the plurality of shared memory servers 120 create and allocate a shared memory buffer segment according to a request for creation and allocation of the shared memory buffer, respectively, to provide a virtual shared memory buffer connecting the respective shared memory buffer segments. can Here, the shared memory buffer may be directly accessed by the distributed processing unit 110 through the remote direct memory access support network 130 .

In this case, the shared memory buffer may be synchronized with the local shared memory area of the distributed processing devices 110 . That is, synchronization is possible through RDMA read/write.

In this case, the shared memory server 120 may provide an accumulation operation function between the shared memory buffer segments.

The remote direct memory access support network 130 is a network that provides communication between a plurality of distributed processing devices 110 and a plurality of shared memory servers 120 , and the distributed processing devices 110 are shared memory servers. Provides a function that directly accesses the memory of 120.

That is, in a high-performance computing cluster system environment connected by a high-performance network 130 supporting remote direct memory access, the distributed processing units 110 combine the physical memory segments of a plurality of shared memory servers 120 to create a virtual continuous By providing direct access to the shared memory buffer, data sharing between distributed processing units can be accelerated and efficiency can be increased.

As a representative distributed processing method that shares distributed processing data in the form of such a shared memory, it can be used for asynchronous deep learning training. The asynchronous deep learning training method is one of the data parallel deep learning learning methods. Multiple deep learning processes perform learning by dividing the training data, and the contents learned during learning are shared with other deep learning processes through a shared data buffer. A shared learning method. In the asynchronous deep learning training method, each deep learning distributed processing process updates the parameters asynchronously without synchronizing the deep learning parameters (general names of weights and feature values to be trained in deep learning training) with other processes. This method is suitable for the shared memory structure proposed by the present invention.

In addition, the method proposed in the present invention is somewhat similar to the parameter server, but unlike the parameter server method that receives a gradient from the deep learning distributed processes and updates the deep learning parameters directly by actively calculating the weight parameters, the distributed processing devices are remotely and directly The memory access function allows direct reading and writing of shared memory buffers without intervention of the shared memory server.

In addition, if only the memory on one single memory server is used as shared memory, scalability is limited, but flexible scalability can be provided even for large-scale distributed processing by configuring a shared memory server cluster.

FIG. 2 is a block diagram illustrating an example of the distributed processing apparatus 110 through remote direct memory access shown in FIG. 1 .

Referring to FIG. 2 , the distributed processing apparatus 110 through remote direct memory access according to an embodiment of the present invention includes a control unit 210 , a communication unit 220 , a memory 230 , an operation processing unit 240 , and a shared memory. and a server access support unit 250 and the like.

In detail, the control unit 210 controls the distributed processing process through remote direct memory access as a kind of central processing unit. That is, the controller 210 may provide various functions by controlling the memory 230 , the arithmetic processing unit 240 , and the shared memory server access support unit 250 .

Here, the controller 210 may include all kinds of devices capable of processing data, such as a processor. Here, the 'processor' may refer to a data processing device embedded in hardware, for example, having a physically structured circuit to perform a function expressed as a code or an instruction included in a program. As an example of the data processing apparatus embedded in the hardware as described above, a microprocessor, a central processing unit (CPU), a processor core, a multiprocessor, an application-specific integrated (ASIC) circuit) and a processing device such as a field programmable gate array (FPGA), but the scope of the present invention is not limited thereto.

Communication unit 220 through the RDMA support network (see 130 in Fig. 1) through the remote direct memory access through the distributed processing unit 110 and the shared memory server (see 120 in Fig. 1) a communication interface necessary to transmit a transmission/reception signal to provide.

Here, the communication unit 220 may be a device including hardware and software necessary for transmitting and receiving signals such as control signals or data signals through wired/wireless connection with other network devices.

In this case, the communication unit 220 may directly access remote memories of the shared memory servers (see 120 of FIG. 1 ) through the RDMA support network (see 130 of FIG. 1 ) to read and write data.

The memory 230 performs a function of temporarily or permanently storing data processed by the controller 210 . Here, the memory 230 may include magnetic storage media or flash storage media, but the scope of the present invention is not limited thereto.

In this case, the memory 230 may allocate the same size as the shared memory buffer configured from remote memories of the shared memory servers (see 120 of FIG. 1 ) to the local shared memory area, and may be synchronized with the shared memory buffer.

Accordingly, the distributed processing unit 110 may share memory with other distributed processing units by performing an operation on the local shared memory area and synchronizing it with the shared memory buffer.

The arithmetic processing unit 240 performs an operation given to the distributed processing unit 110 in the distributed processing framework.

In this case, the operation processing unit 240 may perform an operation on the local shared memory area.

In this case, the operation processing unit 240 may explicitly request and perform synchronization of the local shared memory area and the shared memory buffer to the shared memory server access support unit 250 through an API (Application Programmable Interface).

In this case, the operation processing unit 240 may request and perform an accumulation operation on the plurality of shared memory buffers to the shared memory server access support unit 250 through an application programmable interface (API).

The shared memory server access support unit 250 supports access through RDMA read/write with the shared memory server (see 120 of FIG. 1 ), and supports synchronization with the shared memory buffer.

In this case, the shared memory server access support unit 250 may acquire information on the shared memory servers (see 120 of FIG. 1 ) constituting the shared memory server cluster by registering the shared memory server cluster. Here, the registration of the shared memory server cluster establishes a connection with the corresponding shared memory servers (see 120 in FIG. 1) using the access information of the shared memory servers (see 120 in FIG. 1) input by the user, and the shared memory This may refer to an initialization process in which a buffer segment is created and allocated. In addition, the shared memory server access information may include IP address and port number information. In particular, when registering the shared memory server cluster, all distributed processing units 110 may register the same order of all shared memory servers (see 120 of FIG. 1 ).

In this case, the shared memory server access support unit 250 may calculate the size of the shared memory buffer segments constituting the shared memory buffer for each shared memory server (see 120 of FIG. 1 ) registered in the shared memory server cluster. For example, when five shared memory servers are included in one shared memory server cluster and the size of the shared memory buffer is configured to be 5 GB, it can be divided into five shared memory buffer segments having a size of 1 GB. Here, the size of each shared memory buffer segment may not be the same.

At this time, the shared memory server access support unit 250 transmits a request to the shared memory servers (see 120 of FIG. 1 ) the size information of the calculated shared memory server segments and the creation and allocation of the shared memory buffer segment to create the shared memory buffer. configurable.

In this case, the shared memory server access support unit 250 may obtain an access right to the shared memory buffer configured by another distributed processing device to replace the creation and allocation of the shared memory buffer.

In this case, the shared memory server access support unit 250 may transmit the shared memory buffer creation key together to the shared memory servers (see 120 of FIG. 1 ) to configure the shared memory buffer. Here, the shared memory buffer creation key may be used for the purpose of preventing duplicate creation of the same shared memory buffer, checking whether a request is a valid request, or specifying a shared memory buffer segment.

At this time, when the shared memory server access support unit 250 creates and allocates shared memory buffer segments for all shared memory servers (see 120 of FIG. 1 ), the shared memory buffer is locally shared in the memory 230 with the same size as the shared memory buffer. A memory area can be allocated. Here, the local shared memory area is allocated to real physical memory, and address information (eg, virtual address) is returned as the local shared memory area is allocated.

In this case, the shared memory server access support unit 250 may synchronize the local shared memory area with the shared memory buffer. Here, synchronization may be performed using a memory mapping table.

In this case, when the data of the local shared memory area is changed by the operation processing unit 240 , the shared memory server access support unit 250 may synchronize the data of the local shared memory area by copying it to the shared memory buffer through RDMA. Alternatively, only the changed data can be copied and synchronized.

In this case, the shared memory server access support unit 250 may synchronize by copying the data of the shared memory buffer to the local shared memory area through RDMA when the data of the shared memory buffer is changed by the operation of another distributed processing device. Alternatively, only the changed data can be copied and synchronized.

In this case, the shared memory server access support unit 250 may request the shared memory servers (see 120 of FIG. 1 ) to release and delete the shared memory buffer segments when the use of the shared memory buffer is terminated.

At this time, when all shared memory buffer segments are released and deleted, the shared memory server access support unit 250 terminates the connection with the shared memory servers (see 120 in FIG. 1 ), deregisters the shared memory server cluster, and information You can end the use of the shared memory buffer by deleting .

In this case, the shared memory server access support unit 250 may provide a data accumulation operation function for a plurality of shared memory buffers. For example, when performing an accumulation operation on the first shared memory buffer and the second shared memory buffer, data synchronization is performed on the first shared memory buffer, and the first shared memory server (see 120 in FIG. 1 ) is An accumulation operation from the shared buffer segments to the second shared buffer segments is requested, and when the accumulation operation is completed in all shared memory servers (see 120 of FIG. 1 ), the result may be returned. Each of the shared memory servers (see 120 of FIG. 1 ) may lock the second shared memory buffer segment for an accumulation operation, and may perform an accumulation operation from the first shared memory buffer segment to the second shared memory buffer segment.

The memory mapping table management unit 260 manages a memory mapping table between the local shared memory area and the shared memory buffer.

In this case, the memory mapping table management unit 260 may directly store and manage the memory mapping table including the storage, but may store and manage the memory mapping table in a separate storage or memory 230 .

In this case, when the shared memory buffer is created, the memory mapping table management unit 260 may create a memory mapping table between the shared memory buffer and the local shared memory area.

In this case, the memory mapping table management unit 260 may delete the memory mapping table when the use of the shared memory buffer is terminated.

3 is a block diagram illustrating an example of the shared memory server 120 shown in FIG.

Referring to FIG. 3 , the shared memory server 120 according to an embodiment of the present invention includes a control unit 310 , a communication unit 320 , a memory 330 , and a shared memory management unit 340 .

In detail, the control unit 310 controls a distributed processing process through remote direct memory access as a kind of central processing unit. That is, the controller 310 may provide various functions by controlling the memory 330 and the shared memory management unit 340 .

Here, the controller 310 may include all kinds of devices capable of processing data, such as a processor. Here, the 'processor' may refer to a data processing device embedded in hardware, for example, having a physically structured circuit to perform a function expressed as a code or an instruction included in a program. As an example of the data processing apparatus embedded in the hardware as described above, a microprocessor, a central processing unit (CPU), a processor core, a multiprocessor, an application-specific integrated (ASIC) circuit) and a processing device such as a field programmable gate array (FPGA), but the scope of the present invention is not limited thereto.

The communication unit 320 is a communication interface necessary for transmitting a transmission/reception signal between the shared memory server 120 and the distributed processing devices (refer to 110 of FIG. 1 ) through the remote direct memory access through the RDMA support network (see 130 of FIG. 1 ). provides

Here, the communication unit 320 may be a device including hardware and software necessary for transmitting and receiving signals such as control signals or data signals through wired/wireless connection with other network devices.

In this case, the communication unit 320 allows the distributed processing units (refer to 110 in FIG. 1) to directly access the memory 330 through remote direct memory access through the RDMA support network (see 130 in FIG. 1) to read and write data. can support you to

The memory 330 performs a function of temporarily or permanently storing data processed by the controller 310 . Here, the memory 330 may include magnetic storage media or flash storage media, but the scope of the present invention is not limited thereto.

In this case, all or part of the memory 330 may be allocated as a shared memory buffer segment, and may constitute a shared memory buffer together with the shared memory buffer segments of other shared memory servers. Here, the shared memory buffer is a virtual memory buffer that connects the shared memory buffer segments, and the substance is an area of the shared memory buffer segments allocated to the memory 330 of each shared memory server 120 .

In this case, the shared memory buffer segment allocated to the memory 330 may synchronize the changed data as data in the local shared memory area of the distributed processing unit (refer to 110 of FIG. 1 ) is changed.

The shared memory management unit 340 allocates a shared memory buffer segment to configure the shared memory buffer and manages it.

At this time, when the shared memory management unit 340 receives a request for creation and allocation of the shared memory buffer from the distributed processing unit (see 110 of FIG. 1 ), the shared memory buffer segment is configured in the memory 330 by the size of the given shared memory buffer segment. You can create and allocate to return allocation information.

At this time, the shared memory management unit 340 receives the shared memory buffer creation key in response to the shared memory buffer segment creation request, and creates and allocates the shared memory buffer segment only when the received shared memory buffer creation key is not already in use. Can return allocation information.

At this time, the shared memory management unit 340 receives the shared memory buffer generation key in response to the request for access of the shared memory buffer segment, and when the received shared memory buffer generation key matches information of the shared memory buffer segment requesting access You can allow access to the corresponding shared memory buffer segment. For example, the shared memory buffer creation key may mean the size of the shared memory buffer segment.

In this case, the shared memory management unit 340 may perform a data accumulation operation between the shared memory buffer segments. For example, accumulating data from the first shared memory buffer segment to the second shared memory buffer segment may be performed by locking the second shared memory buffer segment and accumulating the data of the first shared memory buffer segment into the second shared memory buffer segment. have. In addition, when the data accumulation operation is completed, a result indicating the completion of the operation may be returned.

4 is a diagram illustrating a method of configuring a shared memory buffer according to an embodiment of the present invention.

4, in the method of configuring a shared memory buffer according to an embodiment of the present invention, the master distributed processing unit 410 that leads the creation and allocation of the shared memory buffer includes the shared memory servers 440, 450, 460 and 470) to request creation and allocation of shared memory buffer segments for constituting a shared memory buffer.

At this time, the shared memory server 1 (440) creates and allocates the shared memory buffer segment 1 (441) and returns the information, and the shared memory server 2 (450) creates and allocates the shared memory buffer segment 2 (451). Returning the information, shared memory server 3 (460) creates and allocates shared memory buffer segment 3 (461) to return the information, and shared memory server 4 (470) creates shared memory buffer segment 4 (471). It can be created and assigned to return that information.

At this time, when the shared memory buffer segments 441 , 451 , 461 and 471 are created and allocated in each of the shared memory servers 440 , 450 , 460 and 470 , they are connected to form a virtual shared memory buffer 430 . can do. That is, the substance of the shared memory buffer 430 is the shared memory buffer segments 441 , 451 , 461 and 471 allocated to each of the shared memory servers 440 , 450 , 460 and 470 . Accordingly, data input/output to the shared memory buffer 430 means data input/output to the shared memory buffer segments 441 , 451 , 461 and 471 .

At this time, other distributed processing units belonging to the same distributed processing framework as the master distributed processing unit 410 are classified as slave distributed processing units 420 , and the slave distributed processing unit 420 is controlled by the master distributed processing unit 410 . By acquiring information of the configured shared memory buffer, the same shared memory buffer can be used.

In this case, the master distributed processing unit 410 and the slave distributed processing unit 420 may allocate local shared memory areas 411 and 421 having the same size as the shared memory buffer for their own memories, respectively. In addition, distributed processing in the distributed processing framework may be performed by performing input/output to the local shared memory regions 411 and 421 .

At this time, each of the local shared memory regions 411 and 421 is maintained in synchronization with the shared memory buffer 430 , and distributed processing data is shared between the distributed processing units 410 and 420 through the shared memory buffer 430 . can do. In particular, the distributed processing units 410 and 420 may perform direct input/output to and from the shared memory buffer 430 through RDMA to perform synchronization between the local shared memory areas 411 and 421 and the shared memory buffer 430 . .

5 is a flowchart illustrating a distributed processing method through remote direct memory access according to an embodiment of the present invention.

Referring to FIG. 5 , in the distributed processing method through remote direct memory access according to an embodiment of the present invention, a distributed processing device (refer to 110 in FIG. 1 ) through remote direct memory access registers a shared memory server cluster ( S501). The shared memory server cluster may be composed of a plurality of shared memory servers (see 120 of FIG. 1 ).

In addition, in the distributed processing method through remote direct memory access according to an embodiment of the present invention, the distributed processing device (refer to 110 of FIG. 1 ) through remote direct memory access creates and allocates a shared memory buffer (S503).

In addition, in the distributed processing method through remote direct memory access according to an embodiment of the present invention, the distributed processing unit (see 110 in FIG. 1 ) through remote direct memory access performs a given distributed operation and RDMA for the shared memory buffer The distributed processing data is shared through direct data read and write through (S505).

In addition, in the distributed processing method through remote direct memory access according to an embodiment of the present invention, when the distributed processing device (see 110 in FIG. 1 ) through remote direct memory access ends the use of the shared memory buffer, the shared memory The buffer is released and deleted (S507).

In addition, in the distributed processing method through remote direct memory access according to an embodiment of the present invention, the distributed processing device (see 110 of FIG. 1 ) through remote direct memory access deregisters the shared memory server cluster (S509) .

6 is an operation flowchart illustrating an example of the step ( S503 ) of creating and allocating the shared memory buffer shown in FIG. 5 .

Referring to FIG. 6 , the step of creating and allocating the shared memory buffer shown in FIG. 5 ( S503 ) is performed by the distributed processing unit (see 110 of FIG. 1 ) through remote direct memory access, the shared memory buffer segment for each shared memory server The size is calculated (S601).

In addition, in the step (S503) of creating and allocating the shared memory buffer shown in FIG. 5, the distributed processing unit (refer to 110 in FIG. 1) through remote direct memory access is to the shared memory servers (refer to 120 in FIG. 1). It requests creation and allocation of a shared memory buffer segment (S603). Here, a request for creation and allocation of a shared memory buffer segment may be used in the same meaning as a request for creation and allocation of a shared memory buffer.

In addition, in the step (S503) of creating and allocating the shared memory buffer shown in FIG. 5, the distributed processing unit (refer to 110 in FIG. 1) through remote direct memory access is to the shared memory servers (refer to 120 in FIG. 1). When the shared memory buffer segments are created and allocated, information returned is obtained (S605).

In addition, the step (S503) of creating and allocating the shared memory buffer shown in FIG. 5 is the distributed processing unit (refer to 110 in FIG. 1) through remote direct memory access, whether all shared memory buffer segments have been created and allocated. Confirm (S607).

In addition, in the step of creating and allocating the shared memory buffer shown in FIG. 5 ( S503 ), the distributed processing unit (see 110 of FIG. 1 ) through remote direct memory access creates and allocates all shared memory buffer segments to the shared memory When the buffer is configured, a local shared memory area of the same size as the shared memory buffer is allocated to the memory, and a memory mapping table between the shared memory buffer and the local shared memory area is updated (S609).

7 is an operation flowchart illustrating an example of the step of releasing and deleting the shared memory buffer shown in FIG. 5 ( S507 ).

Referring to FIG. 7 , the step of releasing and deleting the shared memory buffer shown in FIG. 5 ( S507 ) is performed by the distributed processing unit (see 110 of FIG. 1 ) through remote direct memory access, and the shared memory servers ( FIG. 1 ). 120) to request release and deletion of the shared memory buffer segment (S701). Here, the request to release and delete the shared memory buffer segment may be used in the same meaning as the request to release and delete the shared memory buffer.

In addition, the step of releasing and deleting the shared memory buffer shown in FIG. 5 ( S507 ) is performed by the distributed processing unit (see 110 in FIG. 1 ) through remote direct memory access to the shared memory servers (see 120 in FIG. 1 ). When the shared memory buffer segments are released and deleted, the returned information is obtained (S703).

In addition, the step of releasing and deleting the shared memory buffer shown in Fig. 5 (S507) is a distributed processing unit (see 110 in Fig. 1) through remote direct memory access, whether all shared memory buffer segments have been released and deleted Confirm (S705).

In addition, in the step of releasing and deleting the shared memory buffer shown in FIG. 5 ( S507 ), the distributed processing unit (see 110 in FIG. 1 ) through remote direct memory access releases and deletes all shared memory buffer segments to the shared memory When the use of the buffer is finished, the allocated local shared memory area is released, and the corresponding memory mapping table is deleted (S707).

8 is an operation illustrating a data accumulation operation method of shared memory buffers according to an embodiment of the present invention.

Referring to FIG. 8 , in the method for accumulating data of shared memory buffers according to an embodiment of the present invention, a data accumulating operation is performed on shared memory buffer segments allocated by each of the shared memory servers 810 , 820 and 830 . done by performing

A first shared memory buffer segment 1 (841) and a second shared memory buffer segment 1 (851) are allocated to the first shared memory server 810, and the first shared memory buffer segment is allocated to the second shared memory server 820. 2 ( 842 ) and a second shared memory buffer segment 2 ( 852 ) are allocated, and an nth shared memory server ( 830 ) has a first shared memory buffer segment n ( 843 ) and a second shared memory buffer segment n ( 853 ) is assigned. In addition, the first shared memory buffer segment 1 841 , the first shared memory buffer segment 2 842 , and the first shared memory buffer segment n 843 , etc. constitute the first shared memory buffer 840 . In addition, the second shared memory buffer segment 1 ( 851 ), the second shared memory buffer segment 2 ( 852 ), and the second shared memory buffer segment n ( 853 ) constitute the second shared memory buffer ( 850 ).

Each of the shared memory servers 810 , 820 and 830 share the data of the first shared memory buffer segment 841 , 842 and 843 of the first shared memory buffer 840 with the second shared memory buffer 850 . By accumulating in the memory buffer segments 851 , 852 , and 853 , a data accumulation operation between the shared memory buffers may be performed.

9 is an operation flowchart illustrating a data accumulation operation method of the shared memory buffers shown in FIG. 8 .

Referring to FIG. 9 , in the method of accumulating data of the shared memory buffers shown in FIG. 8 , the distributed processing unit (see 110 of FIG. 1 ) through remote direct memory access transfers the data of the first shared memory buffer to the local shared memory area. is synchronized with (S901).

In addition, in the method for accumulating data of shared memory buffers shown in FIG. 8 , a distributed processing unit (see 110 in FIG. 1 ) through remote direct memory access provides a first shared memory to shared memory servers (see 120 in FIG. 1 ). A data accumulation operation is requested from the buffer to the second shared memory buffer (S903).

In addition, the data accumulation operation method of the shared memory buffers shown in FIG. 8 is a distributed processing unit (see 110 in FIG. 1 ) through remote direct memory access, a second shared memory from shared memory servers (see 120 in FIG. 1 ) After locking the buffer segment, a result of accumulating the data of the first shared memory buffer segment in the second shared memory buffer segment is received (S905).

In addition, in the data accumulation operation method of the shared memory buffers shown in FIG. 8, the distributed processing unit (refer to 110 in FIG. 1) through remote direct memory access checks whether the accumulation operation for all shared memory buffer segments is completed, The result of the accumulation operation is returned (S907).

The specific implementations described in the present invention are examples and do not limit the scope of the present invention in any way. For brevity of the specification, descriptions of conventional electronic components, control systems, software, and other functional aspects of the systems may be omitted. In addition, the connection or connection members of the lines between the components shown in the drawings exemplarily represent functional connections and/or physical or circuit connections, and in an actual device, various functional connections, physical connections that are replaceable or additional may be referred to as connections, or circuit connections. In addition, unless there is a specific reference such as “essential” or “importantly”, it may not be a necessary component for the application of the present invention.

Therefore, the spirit of the present invention should not be limited to the above-described embodiments, and the scope of the spirit of the present invention is not limited to the scope of the scope of the present invention. will be said to belong to

100: Distributed processing system through remote direct memory access
110: Distributed processing unit through remote direct memory access
120: shared memory server 130: RDMA support network
210: control unit 220: communication unit
230: memory 240: arithmetic processing unit
250: shared memory server access management unit
260: memory mapping table management unit
310: control unit 320: communication unit
330: memory 340: shared memory management unit

Claims (8)

a communication unit for transmitting and receiving data by accessing remote memories provided in a plurality of shared memory servers; and
For a shared memory buffer composed of shared memory buffer segments allocated from the remote memories, a local shared memory area of a size corresponding to that of the shared memory buffer is allocated to a memory, and the shared memory buffer and the local shared memory area are separated Synchronizing shared memory access management unit;
Distributed processing device through remote direct memory access, characterized in that it comprises a. The method according to claim 1,
The shared memory access management unit
Distributed processing unit through remote direct memory access, characterized in that the shared memory buffer and the local shared memory area are synchronized using a memory mapping table. The method according to claim 1,
The shared memory buffer is
Distributed processing unit through remote direct memory access, characterized in that it corresponds to a virtual continuous buffer shared by a plurality of distributed processing units. The method according to claim 1,
The shared memory buffer segments are
Distributed processing device through remote direct memory access, characterized in that corresponding to the physical memory segments of the plurality of shared memory servers. The method according to claim 1,
The shared memory access management unit
Distributed through remote direct memory access, characterized in that calculating the size of the shared memory buffer segments corresponding to each of the shared memory servers, and requesting the shared memory servers to create and allocate the shared memory buffer segments processing unit. The method according to claim 1,
The shared memory access management unit
Distributed processing device through remote direct memory access, characterized in that it performs a data accumulation operation between a plurality of shared memory buffers. a communication unit for transmitting and receiving data to and from a plurality of distributed processing devices through remote direct memory access; and
a memory directly accessible to the distributed processing units;
including,
a shared memory buffer segment is allocated from the memory;
The shared memory buffer segment is characterized in that to configure a shared memory buffer with the shared memory buffer segments of other shared memory servers, the shared memory server. allocating a local shared memory area of a size corresponding to that of the shared memory buffer in a memory for a shared memory buffer composed of shared memory buffer segments allocated from remote memories provided in a plurality of shared memory servers; and
synchronizing the shared memory buffer and the local shared memory area by accessing the remote memories to transmit and receive data;
Distributed processing method through remote direct memory access, characterized in that it comprises a.

Images