WO2024201804A1 - Relay device - Google Patents

Abstract

A relay device (10) relays data to be transferred by remote direct memory access (RDMA) and is provided with: an information acquisition unit that acquires a plurality of pieces of connection information issued when each of a plurality of remote computers (93, 94) is determined to receive the data by RDMA transfer; a transfer destination selection unit that, when a packet containing the entire data or a divided portion of the data is received from a local computer (91), selects, as a transfer destination computer, the remote computer which is the transfer destination of the packet from among the plurality of remote computers; and a packet output unit that causes the packet to include the connection information issued by the transfer destination computer from among the plurality of pieces of connection information acquired by the information acquisition unit and outputs the included packet, thereby performing the RDMA transfer of the packet to the transfer destination computer.

Description

Repeater

The present invention relates to a relay device, such as a network switch, that relays data transferred via RDMA (Remote Direct Memory Access).

RDMA is a communications protocol. In RDMA, a local computer issues a request to establish a connection with a remote computer. The remote computer then transfers data stored in its own memory directly to the remote computer's main memory (Non-Patent Document 1).

In conventional RDMA, the local computer and remote computer are determined on a one-to-one basis when a connection is established. For this reason, after data is output from the local computer, it is not possible to transfer the data to one of multiple candidate transfer destinations. If data output from the local computer could be transferred to one of multiple candidate transfer destinations, there would be an advantage in that, for example, data could be transferred while avoiding busy remote computers among multiple remote computers.

The objective of the present invention is to enable RDMA transfer of data output from a local computer to one of multiple potential transfer destinations.

In order to solve the above problems, the relay device of the present invention is a relay device that relays data transferred by RDMA (Remote Direct Memory Access), and includes an information acquisition unit that acquires multiple pieces of connection information issued by multiple remote computers when each of the multiple remote computers receives the data by RDMA transfer, a transfer destination selection unit that selects a remote computer to which the packet is to be transferred from the multiple remote computers as a transfer destination computer when a packet containing the entire data or a divided part of the data is received from a local computer, and a packet output unit that includes the connection information issued by the destination computer from the multiple connection information acquired by the information acquisition unit in the packet and outputs the packet with the connection information included, thereby transferring the packet to the destination computer by RDMA transfer.

According to the present invention, data output from a local computer can be RDMA transferred to one of multiple potential transfer destinations.

FIG. 1 is a configuration diagram of a system including a relay device according to a first embodiment of the present invention. FIG. 2 is a hardware configuration diagram of the local and remote computers in FIG. FIG. 3 is a block diagram of the relay device of FIG. FIG. 4 is a flowchart for explaining the process executed by the relay device of FIG. FIG. 5 is a diagram illustrating an example of the configuration of the management table. FIG. 6 is a block diagram of a relay device according to the second embodiment. FIG. 7 is a diagram showing an example of the configuration of a PSN table. FIG. 8 is a block diagram of a relay device according to the third embodiment. FIG. 9 is a block diagram of a relay device according to the third embodiment. FIG. 10 is a configuration diagram of a system including a relay device according to the fourth embodiment.

The following describes an embodiment of the present invention and its modified examples with reference to the drawings.

First Embodiment
1, the relay device 10 according to the present embodiment is communicatively connected to first and second local computers (also simply referred to as local) 91 and 92 as transfer sources in RDMA (Remote Direct Memory Access) via a network NW such as the Internet and a router (not shown). The relay device 10 is further communicatively connected by wire or wireless to first and second remote computers (also simply referred to as remote) 93 and 94 as transfer destinations in RDMA. The relay device 10 is configured as a network switch.

The remotes 93 and 94 are configured to be able to execute the same processing, realizing processing redundancy. The relay device 10 transmits data to be processed from the local 91 or 92 to either the remote 93 or 94. The data to be processed is transmitted in a packet including the data. If the data to be processed is long, it is divided into multiple pieces of data and transmitted in packets including each of the divided pieces of data. The remotes 93 and 94 may be devices that perform distributed processing of big data transmitted from the locals 91 and 92. In this case, the data transmitted by the locals 91 and 92 may be transmitted to either the remote 93 or 94.

The locals 91 and 92 and the remotes 93 and 94 in FIG. 1 are also collectively referred to as computers 90. Each of the locals 91 and 92 and the remotes 93 and 94, that is, each computer 90, has the configuration shown in FIG. 2.

As shown in FIG. 2, the computer 90 comprises a CPU (Central Processing Unit) 90A, a main memory 90B, and a non-volatile storage device 90C that stores an OS (Operating System) executed by the CPU 90A, user applications, and other programs. The computer 90 further comprises a communication interface 90D that is connected to the outside. The communication interface 90D includes, for example, an R-NIC (RDMA-Network Interface Card) 90E as an HCA (Host Channel Adapter) for realizing RDMA transfers. The communication interface 90D may include multiple R-NICs 90E.

The R-NIC90E has multiple QPs (Queue Pairs), each consisting of an SQ (Send Queue) and an RQ (Receive Queue), that are used when performing RDMA transfers. The communication unit for RDMA transfers is a communication request called a WR (Work Request), and the data to be communicated is stacked in the SQ/RQ as a WQE (Work Queue Element).

The R-NIC90E also has a Completion Queue (CQ) corresponding to each SQ/RQ, where the WQEs stacked in the SQ/RQ are stacked as Completion Queue Entries (CQEs) when a WR between QPs is completed.

As shown in FIG. 3, the relay device 10 includes an information acquisition unit 11, a forwarding selection unit 12, a packet output unit 13, and a memory unit 14. The above units 11 to 14 are configured as a whole in at least one of an ASIC (Application Specific Integrated Circuit) and an FPGA (Field-Programmable Gate Array). In other words, the entire units 11 to 14 may be configured in either an ASIC or an FPGA, or a portion of the units 11 to 14 may be configured in either an ASIC or an FPGA, and the remainder of the units 11 to 14 may be configured in the other of an ASIC or an FPGA.

The operation of each of the above-mentioned units 11 to 14 of the relay device 10 will be explained with reference to FIG. 4 along with the flow of RDMA transfer. In the following explanation, the RDMA transfer will be explained as a transfer in which data prepared in the main memory 90B of the local 91 is written to the main memory 90B of either the remote 93 or 94.

As shown in FIG. 4, first, the CPU 90A of the local 91 issues a request for establishing a connection for RDMA transfer and establishes a connection with the relay device 10 (step S11). After that, the information acquisition unit 11 of the relay device 10 establishes connections with each of the remotes R1 and R2 (steps S12 and S13).

When a connection is established, each CPU 90A provided in the local 91 and the remotes 93 and 94 determines the memory area for RDMA transfer in the main memory 90B (the local side is the memory area that stores the transmitted data, and the remote side is the memory area that stores the received data). Each CPU 90A also determines which of the multiple QPs of the R-NIC 90E to use. Each CPU 90A generates connection information including the results of these determinations and transmits it to the relay device 10. The transmitted connection information is acquired by the information acquisition unit 11 of the relay device 10.

The connection information includes, for example, the following information:
(1) QP number: A number that identifies the QP used in this RDMA transfer.
(2) Information about the memory area used in this RDMA transfer.
Key: A key for accessing a memory area (address) in the main memory where data to be transferred via RDMA is stored or read.
addr: address of the registered memory area; buffer size: size of the registered memory area. (3) Global ID and local ID of the R-NIC used for RDMA, and global ID and local ID of the computer having this R-NIC.

The information acquisition unit 11 of the relay device 10 registers the acquired connection information of the local 91, the connection information of the remote 93, and the connection information of the remote 94 in the management table (see FIG. 5) of the storage unit 14 in association with each other (step S14). In this embodiment, data to be RDMA transferred is transferred to one of multiple transfer destination candidates (here, remotes 93 and 94). The management table is provided to manage the data transfer destination candidates. As shown in FIG. 5, the management table registers the connection information of the transfer source (such as the local 91) and each connection information of the transfer destination candidates (remote 93 and 94) in association with each other for each unit of RDMA transfer. Note that the connection information may further include information on the application that manages the connection.

The information acquisition unit 11 of the relay device 10 returns a registration notification of the connection information to the local 91 (step S15).

The local 91 receives the registration notification of the connection information and performs an RDMA transfer (step S16). That is, the R-NIC 90E of the local 91 reads the transmission data from the main memory 90B, generates a packet including this transmission data in the payload, and transmits it to the relay device 10. At this time, the R-NIC 90E includes the connection information on the local 91 side in the packet. The position of the connection information in the packet is not limited to the header and can be any position.

The packet from the local 91 is received by the forwarding destination selection unit 12 of the relay device 10. The forwarding destination selection unit 12 selects the forwarding destination of the received packet (step S17).

An example of step S17 is described below. First, the destination selection unit 12 extracts the connection information of the source (local 91) from the header of the first packet, and refers to the management table (Figure 5) based on the extracted connection information of the source. The destination selection unit 12 acquires from the management table a number of pieces of connection information corresponding to the connection information of the source. The destination selection unit 12 recognizes a number of remotes, here remotes 93 and 94, as destination candidates, each of which is identified by a global ID or the like contained in the acquired number of connection information. The destination selection unit 12 then selects one of the recognized destination candidates as the destination.

The method of selecting the transfer destination is arbitrary. The transfer destination is selected, for example, taking into consideration the communication status between the remotes 93 and 94 and the relay device 10. The communication status includes the network status and the busy state of the remotes. For example, the transfer destination selection unit 12 measures the time required for communication with the remote 93 (i.e., the time from the transmission timing of the specified data to the reception timing of the response) and the time required for communication with the remote 94, and selects the remote with the shortest time of the two measured times as the transfer destination. Any communication is adopted as the communication. For example, communication for establishing a connection may be adopted, or communication performed exclusively to grasp the communication status may be adopted. In this way, it is preferable that the remote with a communication status better than a predetermined standard is selected as the transfer destination. As another example, the transfer destination may be selected randomly. As another example, the transfer destination selection unit 12 may count the number of times RDMA communication has been performed for each remote, and the remote with the smallest count value may be the transfer destination. The transfer destination selection unit 12 may also select the remotes in order. Also, the remote may be selected by weighting each connection information in the management table and then selecting the connection information. In this way, load balancing may be performed when selecting the transfer destination.

After step S17, the packet output unit 13 writes the connection information of the destination selected by the destination selection unit 12 into the header of the packet (step S18). Here, it is assumed that the remote 93 has been determined as the destination. In other words, the packet output unit 13 writes the connection information of the remote 93 registered in the management table, which corresponds to the connection information sent this time, into the header.

Then, the packet output unit 13 outputs the packet after writing. The output packet is RDMA transferred to the remote 93 according to the connection information (step S19).

The data transferred by RDMA from the local 91 may be divided into multiple packets and sent. In this case, steps S16 to S19 are repeated the number of times equal to the number of packets. However, in step S17, the same destination as the first packet is selected. This causes a series of multiple packets to be transferred to the same remote, and even the same R-NIC. Whether or not the data is being sent in multiple packets can be determined from the data length included in the packet header, etc.

The R-NIC 90E of the remote 93 references the connection information in the packet's header and performs processing for RDMA transfer. As a result, the packet's payload data is stored as a WQE in the RQ of the QP (specified by (1) above) provided in the R-NIC 90E (specified by (3) above), and then transferred to the memory area of the main memory 90B (specified by (2) above). After transfer, the WQE is stored as a CQE in the CQ corresponding to the RQ, and the packet stored in the RQ is released.

Then, the remote 93 sends a response to the local 91, which is the source of the transfer, via the relay device 10 to notify that the transfer has ended (steps S20 and S21).

Then, the local 91 and the relay device 10 disconnect the connection (step S22). In response to this disconnection, the relay device 10 deletes the record including the connection information used this time from the management table (step S23), and instructs the remote devices 93 and 94 to release the reserved memory area, etc. (steps S24 and S25).

The above operations are also performed for local 92.

In this embodiment, in the RDMA transfer of one data item, both remotes 93 and 94 issue connection information, but the connection information generated by the remote that was not selected as the transfer destination is not used. In other words, the connection information is information issued by each of remotes 93 and 94 when the remotes 93 and 94 receive the data to be transferred by RDMA transfer. The information acquisition unit 11 acquires such connection information. Then, when the transfer destination selection unit 12 receives a packet containing the entire data to be transferred or a divided part of the data from the local 91 or 92, it selects the remote to which the packet is to be transferred from the remotes 93 and 94 as candidate transfer destinations as the transfer destination remote. Then, the packet output unit 13 includes the connection information issued by the transfer destination remote among the multiple connection information items acquired by the information acquisition unit 11 in the header of the packet and outputs the included packet. The connection information included in the header is information necessary to realize the RDMA transfer to the transfer destination remote (more specifically, information specifying the transfer destination of the data), so the packet output from the packet output unit 13 is RDMA transferred to the transfer destination remote. With this configuration, the data transfer destination is determined by the relay device 10, so data output from the local 91 or 92 can be RDMA transferred to one of multiple transfer destination candidates.

In particular, by selecting as the transfer destination a remote with a better (e.g., the best) communication status with the relay device 10 than a predetermined standard, an advantage is obtained, for example, that data is transferred while avoiding a busy remote among multiple remotes. This provides low latency and high throughput. In addition, the transfer destination selection unit 12 selects the transfer destination remote by load balancing, thereby achieving load balancing of the transfer destination for RDMA transfer.

Also, because the relay device 10 is a network switch, the load on the R-NIC 90E or CPU 90A is prevented from being placed on the selection of the destination. Furthermore, the information acquisition unit 11, the destination selection unit 12, and the packet output unit 13 are configured as an ASIC or an FPGA as a whole, which results in lower latency, higher throughput, and higher power efficiency than if the relay device 10 were a server computer.

Second Embodiment
A PSN (Packet Sequence Number) may be managed for each QP in each of the locals 91 and 92 and the remotes 93 and 94. The PSN may be the same as that used in conventional RDMA. The PSN is incremented each time a packet is transmitted or received. When one piece of data is transferred using multiple packets, the PSN of each packet becomes a consecutive number. When multiple RDMA transfers are performed with multiple pieces of data as the target, the PSN becomes a cumulative value. The header of a packet of data transferred from a QP is given the PSN of each QP of the source and destination.

In this embodiment, the relay device 10 also manages each PSN for each QP of local 91 and 92 and remote 93 and 94. In such a case, if there is one local and multiple remotes as transfer destination candidates, the increase in the local PSN = the sum of the increase in each of the multiple remote PSNs. However, there are problems as described below. Note that the following issues are based on the premise that data to be transferred via RDMA is divided into multiple packets and transferred.

(Question 1)
Consider a case where the relay device 10 manages the PSNs by incrementing the PSNs of the locals 91 and 92 and the remotes 93 and 94 each time a packet arrives. In this case, assume that the local 91 transmits different RDMA transfer packets to both the remotes 93 and 94, and each response packet is returned to the relay device 10. If the PSN of the local 91 side counted by the relay device 10 is incremented when a response is received, when the responses arrive at the relay device 10 at the same time, the PSNs of the local 91 side related to each RDMA transfer are integrated, and the PSN does not become a correct value. As a result, the PSN included in the header of the next packet may not match the PSN managed by the relay device 10, resulting in an error.

(Question 2)
The remotes 93 and 94 may count the PSN each time they receive a packet, and include the count value at the time of response in the response packet. In this case, if another RDMA transfer packet arrives at the remote before the response (ACK) corresponding to the last packet, the remote behaves as follows. The following "write req" corresponds to the above RDMA transferred packet.

write req #1 first (PSN=k)
write req #1 middle (PSN=k+1)
...
write req #1 middle (PSN=k+n-2)
write req #1 last (PSN=k+n-1): This is the last packet.
write req #2 first (PSN=k+n)
write req #2 middle (PSN=k+n+1)
ack #1 (PSN=k+n+1)

The local device receiving the response expects PSN = k + n - 1, but in reality PSN = k + n + 1, so a transmission error is detected.

(Question 3)
When a packet retransmission occurs, the PSN is rolled back in the local 91, etc. If the relay device 10 manages the PSNs by incrementing the PSNs of the locals 91 and 92 and the remotes 93 and 94 every time a packet arrives, the above formula will not hold true if a rollback occurs, and normal packet transfer will not be possible.

(Solution)
In order to solve the above-mentioned problems, in this embodiment, a PSN table is provided in the storage unit 14 as shown in FIG.

The configuration example of the PSN table is shown in FIG. 7. In FIG. 7, the local ID is information for identifying the local, and here, the same value as the code (91 or 92) attached to the local is used. The state can be "Sending", which indicates that data is being sent in the middle of being sent when multiple packets are sent, "Acked", which indicates that a response has been received, or "Idle", which indicates that RDMA transfer is not being performed. The transfer destination remote ID is information for identifying the remote selected as the transfer destination, and here, the same value as the code (93 or 94) attached to the remote is used. The local start PSN is the PSN of the local side when the first packet is transferred, and the local end PSN is the PSN of the local side when the last packet is transferred. The remote start PSN is the PSN of the remote side when the first packet is transferred, and the remote end PSN is the PSN of the remote side when the last packet is transferred. In the initial state, the PSN table has multiple rows of information storage areas for each local, and each is set to "Idle".

The destination selection unit 12 compares the PSN contained in the header of a packet that the relay device 10 receives from the local 91 or 92 with each range from the local start PSN to the local end PSN in the PSN table. If the PSN contained in the header does not fall within any of the ranges, the packet is usually the first packet of multiple packets when a single piece of data is divided and transferred. In this case, the destination selection unit 12 selects the destination on the condition that the header of the packet contains information indicating that the packet is the first, and obtains the above information based on the information in the header of the packet, etc.

The transfer destination selection unit 12 sets the PSN included in the header of the first packet as the local start PSN. The transfer destination selection unit 12 can obtain the local end PSN by dividing the data length included in the header of the first packet by the data length of one packet and adding the obtained value to the local start PSN. Furthermore, when selecting a transfer destination, the transfer destination selection unit 12 communicates with the selected remote and obtains the current PSN from the remote. The transfer destination selection unit 12 adds 1 to the obtained PSN and sets the result as the remote start PSN. The transfer destination selection unit 12 adds the value added to the local start PSN to the remote start PSN and sets the added value as the remote start PSN.

The destination selection unit 12 registers each piece of acquired information in the "Idle" column. If there is no "Idle", the destination selection unit 12 overwrites and registers each piece of information in the oldest "Acked" (for example, the one with the smallest PSN value). The state at this time is "Sending". If there is no "Idle" or "Acked", the destination selection unit 12 drops the packet. The destination selection unit 12 also drops the packet if the packet header does not contain information indicating that the packet is the first.

If the PSN included in the packet header falls within any of the ranges from the local start PSN to the local end PSN, the destination selection unit 12 determines that the remote identified by the destination remote ID corresponding to that range is the destination of the packet. The destination selection unit 12 then adds the value obtained by subtracting the local start PSN from the PSN included in the header to the remote start PSN. The destination selection unit 12 writes this added value in the packet header as the PSN of the remote side. This ensures that when the packet is sent to the remote side by the packet output unit 13, the PSN of the packet and the PSN of the remote side will be consistent.

The destination selection unit 12 compares the PSN contained in the header of a packet that the relay device 10 receives from the remote 93 or 94 with each range from the remote start PSN to the remote end PSN in the PSN table. If the PSN contained in the header is not within any of the ranges, the destination selection unit 12 drops the packet. If the PSN is within any of the ranges, the destination selection unit 12 adds a value obtained by subtracting the remote start PSN from the PSN contained in the header to the local start PSN. The destination selection unit 12 writes this added value in the header of the packet as the local side PSN. As a result, when the packet is sent to the local side by the packet output unit 13, the PSN of the packet and the local PSN will be consistent.

As described above, in this embodiment, the correct range of PSNs (start PSN to end PSN) contained in a packet is determined for each RDMA transfer by the PSN table, so the inconveniences described in issues 1 and 2 above do not occur. Also, because each packet is compared with the PSN range, there is no problem even if a rollback occurs. Note that the number of rows in the PSN table should be, for example, at least twice as many as the maximum number of anticipated RDMA transfers.

As described above, a packet includes a PSN that specifies the number of packets sent by the local 91 or the like, and the packet output unit 13 outputs a packet if the PSN included in the packet falls within any of the PSN ranges (PSN table) defined for each RDMA transfer. This allows a packet containing the correct PSN to be output. It also supports rollbacks and the like.

Furthermore, for a non-leading packet that is a packet that contains a portion of the data to be transferred and is not the leading packet, the destination selection unit 12 refers to a PSN table that indicates the correspondence between each of the ranges and the remote that is the destination, and when the PSN included in the non-leading packet falls within any of the ranges, selects the remote that corresponds to that range as the destination. This causes the non-leading packet to be transferred to the same destination as the leading packet.

Third Embodiment
As shown in the management table in Fig. 8, the connection information issued by the remote may be issued in units of QP. In this case, the transfer destination selection unit 12 selects the QP of the transfer destination when selecting the transfer destination remote, and the packet output unit 13 includes in the packet the connection information issued by the transfer destination remote for the QP selected by the transfer destination selection unit 12. When the QP is associated with a virtual environment (container, virtual machine), the relay device 10 functions as a switch or load balancer for the virtual environment.

Also, as shown in the management table of FIG. 9, the connection information issued by the remote may be issued in units of memory area (e.g., the key and addr in (2) above) of the main memory 90B to which the data is to be transferred. In this case, the transfer destination selection unit 12 selects the memory area of the transfer destination when selecting the transfer destination remote, and the packet output unit 13 includes in the packet the connection information issued by the transfer destination remote for the memory area selected by the transfer destination selection unit 12. When using a technology in which the RDMA access destination is the device memory, the relay device 10 functions as a switch or load balancer for the device memory area.

Since the QP and other parameters are set for each application that executes RDMA, this embodiment makes it possible to switch between applications and even load balance from outside the remote. Furthermore, load balancing ensures the fairness of the switched QPs, and the bandwidth of each application is approximately the same. With this embodiment as the main focus, the selection targets by the destination selection unit 12 may be multiple QPs or memory areas within one remote.

Fourth Embodiment
As shown in FIG. 8, the relay device 10 may configure a large-scale cluster in a fat tree together with relay devices 101 to 103 such as other switches. The relay device 10 in FIG. 8 also includes the above-mentioned units 11 to 13. In such a case, the packet output unit 13 of the relay device 10 may detect a malfunction of the transmission path (congestion, disconnection, etc.) when transferring a packet from the local 91 or 92 to the remote 93 or 94 that is the RDMA transfer destination. In FIG. 8, the path (dotted line) between the relay device 10 and the relay device 102 is malfunctioning. In such a case, the packet output unit 13 detects the malfunction of the path by a known method, and rewrites the header of the packet so that the packet is transmitted via another transmission path that bypasses the communication path, that is, the transmission path of the relay device 10 → relay device 101 → relay device 103 → remote 93 or 94. The detour may be set in advance for each malfunctioning path, for example.

Conventionally, congestion detection and other functions were performed by the sending side, which placed a load on the sending side. In this embodiment, congestion detection and other functions are performed by the relay device (especially in the case of a switch that uses an ASIC or FPGA), which reduces the load, reduces latency, increases throughput, and increases power efficiency. It also makes it possible to avoid congestion in RDMA connections and balance the network load.

(Modification)
In the above embodiment, the relay device 10 is a switch, but a router or the like may be used as another example of the relay device 10. The hardware configuration of each of the units 11 to 14 of the relay device 10 is arbitrary, and each of the units 11 to 13 may be configured by a processor that executes a program.

The present invention is not limited to the above-described embodiments and modifications. For example, the present invention includes various modifications to the above-described embodiments and modifications that can be understood by a person skilled in the art within the scope of the technical concept of the present invention. The configurations listed in the above-described embodiments and modifications can be combined as appropriate to the extent that there is no contradiction. It is also possible to delete any of the above-described configurations.

(Additional Note)
The following will exemplify configurations as examples of the above-described embodiment and modified examples.
(Appendix 1)
A relay device that relays data transferred by RDMA (Remote Direct Memory Access),
an information acquisition unit that acquires a plurality of pieces of connection information issued by each of a plurality of remote computers when the data is to be received by the RDMA transfer;
a transfer destination selection unit that, when receiving a packet including the entire data or a divided part of the data from a local computer, selects, from the plurality of remote computers, a remote computer to which the packet is to be transferred as a transfer destination computer;
a packet output unit that causes the packet to include connection information issued by the destination computer among the plurality of connection information acquired by the information acquisition unit, and outputs the packet with the connection information included therein, thereby RDMA-transferring the packet to the destination computer;
A relay device comprising:
(Appendix 2)
the transfer destination selection unit selects, from among the plurality of remote computers, a remote computer having a communication status with the relay device that is better than a predetermined standard as the transfer destination computer;
2. The relay device of claim 1.
(Appendix 3)
the relay device is a network switch,
The information acquisition unit, the transfer destination selection unit, and the packet output unit are configured as a whole in at least one of an ASIC (Application Specific Integrated Circuit) and an FPGA (Field-Programmable Gate Array).
3. The relay device according to claim 1 or 2.
(Appendix 4)
the transfer destination selection unit selects the transfer destination computer by load balancing;
4. A relay device according to claim 1.
(Appendix 5)
the packet includes a Packet Sequence Number (PSN) that identifies the number of packets sent by the local computer;
The packet output unit outputs a packet when the PSN contained in the packet is within any of the ranges of PSNs defined for each RDMA transfer.
5. A relay device according to any one of claims 1 to 4.
(Appendix 6)
the packet is a packet including a portion of the data that is divided and is a non-leading packet that is not a leading packet,
the transfer destination selection unit refers to a table showing a correspondence relationship between each range of the PSN and a remote computer to be a transfer destination, and when the PSN included in the non-first packet is within any of the ranges of the PSN, selects the remote computer corresponding to the range as the transfer destination computer;
6. The relay device of claim 5.
(Appendix 7)
Each of the plurality of pieces of connection information is issued in units of a QP (Queue Pair) or a memory area of a data transfer destination,
The transfer destination selection unit selects a QP or memory area to which the data is to be transferred when selecting the transfer destination computer,
The packet output unit includes, in the packet, the connection information issued by the destination computer for the QP or the memory area selected by the destination selection unit.
7. A relay device according to any one of claims 1 to 6.
(Appendix 8)
the packet output unit rewrites a header of the packet so that the packet is transmitted to the destination computer via another transmission path when the transmission path of the packet to the destination computer is out of order;
8. A relay device according to any one of claims 1 to 7.

10... relay device, 11... information acquisition unit, 12... forwarding destination selection unit, 13... packet output unit, 14... storage unit, 90... computer, 90A... CPU, 90B... main memory, 90C... storage device, 90D... communication interface, 90E... R-NIC, 91... first local computer, 92... second local computer, 93... first remote computer, 94... second remote computer.

Claims (8)

1. A relay device that relays data transferred by RDMA (Remote Direct Memory Access),
   an information acquisition unit that acquires a plurality of pieces of connection information issued by each of a plurality of remote computers when the data is to be received by the RDMA transfer;
   a transfer destination selection unit that, when receiving a packet including the entire data or a divided part of the data from a local computer, selects, from the plurality of remote computers, a remote computer to which the packet is to be transferred as a transfer destination computer;
   a packet output unit that causes the packet to include connection information issued by the destination computer among the plurality of pieces of connection information acquired by the information acquisition unit, and outputs the packet with the connection information included therein, thereby RDMA-transferring the packet to the destination computer;
   A relay device comprising:
2. the transfer destination selection unit selects, from among the plurality of remote computers, a remote computer having a communication status with the relay device that is better than a predetermined standard as the transfer destination computer;
   The relay device according to claim 1 .
3. the relay device is a network switch,
   The information acquisition unit, the transfer destination selection unit, and the packet output unit are configured as a whole in at least one of an ASIC (Application Specific Integrated Circuit) and an FPGA (Field-Programmable Gate Array).
   The relay device according to claim 1 .
4. the transfer destination selection unit selects the transfer destination computer by load balancing;
   The relay device according to claim 1 .
5. the packet includes a Packet Sequence Number (PSN) that identifies the number of packets sent by the local computer;
   The packet output unit outputs a packet when the PSN contained in the packet is within any of the ranges of PSNs defined for each RDMA transfer.
   The relay device according to claim 1 .
6. the packet is a packet including a portion of the data that is divided and is a non-leading packet that is not a leading packet,
   the transfer destination selection unit refers to a table showing a correspondence relationship between each range of the PSN and a remote computer to be a transfer destination, and when the PSN included in the non-first packet is within any of the ranges of the PSN, selects the remote computer corresponding to the range as the transfer destination computer;
   The relay device according to claim 5 .
7. Each of the plurality of pieces of connection information is issued in units of a QP (Queue Pair) or a memory area of a data transfer destination,
   The transfer destination selection unit selects a QP or memory area to which the data is to be transferred when selecting the transfer destination computer,
   The packet output unit includes, in the packet, the connection information issued by the destination computer for the QP or the memory area selected by the destination selection unit.
   The relay device according to claim 1 .
8. the packet output unit rewrites a header of the packet so that the packet is transmitted to the destination computer via another transmission path when the transmission path of the packet to the destination computer is out of order;
   The relay device according to claim 1 .

Images