KR20200112422A - Pcie ntb based data share system between multi host and method thereof - Google Patents

Abstract

The present invention relates to a PCIE NTB-based data share system between multiple hosts, which includes: a CPU for processing data transmission or reception to serve as an independent computer that transmits or receives data; at least three hosts each including a memory that stores the data or allocates the data into memory windows as separate address areas; and at least three host adapters based on PCIe NTB each connected to at least three hosts, each equipped with at least two NTB ports, and having at least one of the two NTB ports connected to an NTB port of the host adapter connected to a neighboring host and the other connected to an NTB port of the host adapter connected to another neighboring host, so as to be connected to two different neighboring hosts to have a ring topology structure through the NTB ports, wherein the memory window is shared through the NTB ports of the host adapters formed on the both hosts, so that the data is transmitted and received between the two hosts. Accordingly, multiple host clusters share data at high speed based on the PCIe NTB, thereby implementing a low-cost and high-performance supercomputer.

Description

PCIE NTB BASED DATA SHARE SYSTEM BETWEEN MULTI HOST AND METHOD THEREOF}

The present invention relates to a PCIe NTB-based data sharing system and method between multiple hosts, and in detail, to a switchless network system and method for transmitting and receiving data between multiple hosts based on a shared data area.

The high performance supercomputer (HPC) system consists of a cluster system by connecting a number of computation nodes with an ultra-high performance interconnection network, and this interconnection technology is the performance of a high-performance supercomputer system. It is an important technical factor that determines the stability and stability.

For the self-development of such a high-performance supercomputer, instead of Infiniband/Ethernet, an interconnect technology with high technology dependence, PCIe-based switching technology that can implement low-power and high-performance functions is used as an alternative interconnect technology.

PCIe-based switching technology is a technology used to connect the CPU, memory, and peripherals inside a computer, but there is a limitation in technical support from PCIe switching chip manufacturers, so to use PCIe switching technology as an alternative interconnect technology, PCIe NTB (Non -Transparent Bridge) technology is being proposed.

PCIe NTB is a device that implements direct access through address translation in an address area separated between two independent hosts, and research and development of connections between three or more hosts is required.

US Patent Publication US2014/0075079 (DATA STORAGE DEVICE CONNECTED TO A HOST SYSTEM VIA A PERIPHERAL COMPONENT INTERCONNECT EXPRESS (PCIE) INTERFACE)

The technical problem to be achieved by the present invention is to reduce the cost required to construct a high-performance supercomputer interconnection network by providing a switchless interconnection communication method between three or more multi-hosts using PCIe NTB technology.

The PCIE NTB-based data sharing system between multiple hosts according to an embodiment of the present invention is an independent computer that transmits or receives data, a CPU that processes data transmission or reception, and an address area that stores or separates data. At least three hosts each including memory allocated as a memory window, and each connected to the at least three hosts, each having at least two NTB ports, and one of the at least two NTB ports is connected to the neighboring host. PCIe NTB-based, which is connected to the NTB port of the connected host adapter, and the other is connected to the NTB port of the host adapter connected to another neighboring host, and is connected to two different neighboring hosts in a ring topology structure through NTB ports. Including at least three host adapters of, the memory window is shared through NTB ports of host adapters formed in the two hosts, so that data transmission/reception between the two hosts is performed.

The memory windows are respectively formed for two different neighboring hosts connected through a host adapter of a host on which the memory window is formed, and data between the two hosts are transmitted through memory windows respectively formed for the two different neighboring hosts. It is characterized by sharing.

The memory window includes a memory window for data reception and a memory window for data transmission, and shares data between two hosts through the memory window according to data transmission or reception by the host.

Each of the memory windows formed for the two different neighboring hosts includes a memory window for data reception and a memory window for data transmission, and data between the two hosts is transmitted through the memory window according to data transmission or reception by the host. It is characterized by sharing.

Data transmission/reception between the two hosts is characterized in that it is performed through DMA.

The host transmitting data is characterized in that after transmitting the data through the DMA, an interrupt notifying of the DMA data transmission is generated through the doorbell register, and the host receiving the data checks the received data through the interrupt.

The CPU forms the identification numbers of the at least three hosts as a ring table, and when the host transmits data to another host, the identification number of the host transmitting data and the host to receive the data refer to the ring table. The identification number of is transmitted as information.

The information may further include a memory address of data to be shared and a size of transmission data.

In one example, a scratch pad register is connected between a host transmitting data and a host receiving data, and the host transmitting the data transmits the information to the host receiving the data through the scratch pad register. It is characterized by transmitting.

In a ring topology between host adapters in which the at least three hosts are not directly connected through the host adapter, as a specific host receives data and the information, the identification number of the host that will receive the data included in the information is itself It compares with the identification number of the host, reads the data from the memory window if they match, and forwards the received data to a neighboring host if they do not match.

A method for sharing data between multiple hosts based on PCIE NTB according to an embodiment of the present invention includes at least three hosts and at least three host adapters each connected to the at least three hosts and including at least two NTB ports. In a method for sharing data in a PCIE NTB-based data sharing system between multiple hosts, the step of: a sending host transmits data through the host adapters, and transmits an identification number of a sending host and a destination host identification number of the data as information (S100), the step of receiving data by the host connected to the sending host (S200), the step of the receiving host to determine whether the destination host identification number included in the information matches its own host identification number (Q100), and as a result of performing the determination step (Q100), as the destination host identification number included in the information matches its host identification number, the host that received the data through the memory area copied to the memory window As a result of performing the data verification step (S310) and the determination step (Q100), the destination host identification number included in the information is inconsistent with its own host identification number. It characterized in that it comprises the step of performing any one of the step of forwarding (S320).

According to these features, the PCIe NTB-based data sharing system and method between multiple hosts according to an embodiment of the present invention allows multiple host clusters to share data at high speed based on PCIe NTB without relying on the existing interconnect technology. , It has the effect of implementing a low-cost, high-performance supercomputer.

1 is a block diagram showing a schematic structure of a PCIE NTB-based data sharing system between multiple hosts according to an embodiment of the present invention.
2 is a block diagram showing the structure of a memory window in a data sharing system between multiple hosts based on PCIE NTB according to an embodiment of the present invention.
3 is a block diagram showing a schematic structure of a PCIE NTB-based data sharing system between multiple hosts according to an embodiment of the present invention.
4 is a flowchart illustrating a schematic flow of a method for sharing data between multiple hosts based on PCIE NTB according to an embodiment of the present invention.
5 is a flow chart illustrating in detail a step of a transmitting host transmitting data in a method for sharing data between multiple hosts based on PCIE NTB according to an embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those of ordinary skill in the art can easily implement the embodiments of the present invention. However, the present invention may be implemented in various different forms and is not limited to the embodiments described herein. In the drawings, parts irrelevant to the description are omitted in order to clearly describe the present invention, and similar reference numerals are assigned to similar parts throughout the specification.

1 is a block diagram showing a schematic structure of a PCIE NTB-based multi-host data sharing system according to an embodiment of the present invention, and FIG. 2 is a PCIE NTB-based multi-host data according to an embodiment of the present invention. A block diagram showing the structure of a memory window in a sharing system, and FIG. 3 is a block diagram showing a schematic structure of a PCIE NTB-based data sharing system between multiple hosts according to an embodiment of the present invention, and FIG. A flow chart showing a schematic flow of a method for sharing data between multiple hosts based on PCIE NTB according to an embodiment, and FIG. 5 is a flowchart showing a method for sharing data between multiple hosts based on PCIE NTB according to an embodiment of the present invention. This is a flow chart detailing the steps for transmitting data.

In this embodiment with reference to FIG. 1, the PCIe NTB-based multi-host data sharing system 1 includes first to third hosts 10, 20, 30, which are at least three hosts, and the first to third hosts. It includes first to third host adapters 100, 200, and 300 connected to the hosts 10, 20, and 30, respectively.

The first to third hosts 10, 20, and 30 are independent computers, and each of the hosts 10, 20, and 30 includes first to third CPUs (Central Processing Units) 11, 21, and 31. It includes first to third memories 12, 22 and 32.

The first to third CPUs (11, 21, 31) transmit and receive data to and from other hosts through the first to third host adapters (100, 200, 300) by the first to third hosts (10, 20, 30) Process to do it. In one example, the first CPU 11 processes the first host 10 to transmit and receive data to and from the second host 20 and the third host 30 through the first host adapter 100, respectively, and 2 The CPU 21 processes the second host 20 to transmit and receive data to and from the first host 10 and the third host 30 through the second host adapter 200, respectively, and the third CPU 31 Is processed so that the third host 30 transmits and receives data to and from the first host 10 through the third host adapter 300 and the third NTB 32, and the third host 30 transmits and receives data to and from the third host adapter. 300) and the second NTB 42 to transmit and receive data to and from the second host 20.

As described above, since the first to third CPUs 11, 21, and 31 have a structure that transmits and receives data between hosts through the first to third host adapters 100, 200, and 300, at least three or more hosts are The included multi-host cluster is implemented as a PCIe NTB-based interconnect network, and accordingly, it is possible to implement a PCIe NTB-based HPC system at low cost.

The first to third CPUs 11, 21, 31 of the first to third hosts 10, 20, 30 process data transmission/reception between hosts through the first to third host adapters 100, 200, 300 In detail, the first CPU 11 connects to the third host adapter 300 through the 1-1 NTB port 101 of the first host adapter 100 and connects to the third host 30. ) And share data, and share data with the second host 20 by connecting to the second host adapter 200 through the 1-2 NTB port 102 of the first host adapter 100.

At this time, in an example in which the first to third host adapters 100, 200, and 300 have a ring topology structure as shown in FIG. 1, each host included in the network is another host connected to the host adapter connected to the host Adapters, in detail, through the second host adapter 200 or the third host adapter 300 connected by the first host adapter 100 connected to the first host 10 In this case, as one host adapter is connected to two different host adapters, each host can share data with all other hosts formed in the network.

In an embodiment in which the first to third CPUs 11, 21, and 31 share data between hosts, the first to third CPUs 11, 21, and 31 are provided with memory address areas for hosts other than their own host. It is created, and data is shared between hosts by referring to the created memory address area. In one example, the first to third CPUs 11, 21, 31 are connected to the first to third hosts 10, 20, 30 through the first to third host adapters 100, 200, and 300 in a ring topology. When the structure is initially formed, it creates memory address areas for hosts other than its own host.

The first CPU 11 creates a memory address area for the second host 20 and a memory address area for the third host 30 as separate memory address areas, and stores them in the first memory 12. I can.

In addition, the second CPU 12 creates a memory address area for the first host 10 and a memory address area for the third host 30 as separate memory address areas and stores them in the second memory 12. The third CPU 13 generates a memory address area for the first host 10 and a memory address area for the second host 20 as separate memory address areas and stores them in the third memory 13. .

At this time, the first to third CPUs 11, 21, and 31 allocate a memory address area for a host other than their own host as a memory window for each host, and the first to third memories 12, 22, and 32 Can be saved on. In one example, the first CPU 11 creates a memory window for the third host 30 connected to the 1-1 NTB port 101 of the first host adapter 100 and stores the memory window in the first memory 12. Then, a memory window for the second host 20 connected to the 1-2th NTB port 102 of the first host adapter 100 is generated and stored in the first memory 12.

In this way, as the CPU formed in the host allocates memory windows for other hosts connected to the NTB port of the host adapter and stores them in the memory formed in the host, the first CPU 10 is stored in the first memory 12. Data transmission/reception between the first host 10 and the second host 20 is processed with reference to the memory window, or data transmission/reception between the first host 10 and the third host 30 is processed.

At this time, in an example with reference to FIG. 2, in the PCIe NTB-based multi-host data sharing system 1a, the memory window 120 formed in the first host 10 is for other hosts connected to the host. The memory window is separated and included, and includes a memory window 121 for a second host and a memory window 122 for a third host. Further, the memory windows 121 and 122 for each host may include detailed items such as memory windows 1211 and 1221 for receiving data and memory windows 1212 and 1222 for transmitting data.

As the memory window that shares data with different hosts is formed separately, and has a structure in which the data receiving part and the data transmitting part are separated and formed, PCIe NTB-based interconnect network in a multi-host cluster The reliability of the product is improved.

In addition, the address of the memory window allocated to the first memory 12 is converted through a translation register of the first host adapter 100 to be converted to the first NTB port of the first host adapter 100 ( It is connected to the third memory 32 of the third host 30 connected to 101, or the second memory 22 of the second host 20 connected through the 1-2 NTB port 102, , The memory window and the host adapter become a path to access the memory between the two hosts.

However, in another example with reference to FIG. 3, the first to fourth host adapters 100, 200, 300, and 400 are formed in a ring topology in the same manner that the host adapters shown in FIG. 1 have a ring topology. , Since the first to the first host adapters 100, 200, 300, and 400 each have a structure that is connected only to two different host adapters, the first host 10 and the third host 30 are the first host adapters It has a structure that is not directly connected through the 100 and the third host adapter 300, the second host 20 and the fourth host 40, and the second host adapter 200 and the fourth host adapter 400 It has a structure that is not directly connected through.

Therefore, when there are hosts that are not directly connected through host adapters in a structure including at least four hosts as shown in FIG. 3, an identification number (ID) is assigned to each host to share data between hosts that are not directly connected. Use this to share data.

Or, in another embodiment, the first to fourth host adapters 100, 200, 300, and 400 each assign identification numbers to their hosts 10, 20, 30, and 40, respectively, and 4 Can be stored in the memory (12, 22, 32, 42), and the initial formation of the ring topology structure should not limit the subject that assigns identification numbers to the first to fourth hosts (10, 20, 30, 40). .

In this specification, an embodiment in which each host adapter is implemented in a structure in which each host adapter is connected to another host by having two NTB ports, respectively, is described, but in another example, the host adapter and its NTB port are extended to a different topology. It may be implemented, and in this case, a structure connected by multiple ports may be extended, but is not limited thereto.

In one embodiment, for data sharing between the first host 10 and the third host 30 or between the second host 20 and the fourth host 40, the first host 10 The first memory 12 stores the identification code of the first host 10, the second memory 22 of the second host 20 stores the identification number of the second host 20, and the third host The third memory 32 of 30 stores the identification number of the third host 30, and the fourth memory 42 of the fourth host 40 stores the identification number of the fourth host 40. .

The identification numbers of the first to fourth hosts 10, 20, 30, 40 stored by the first to fourth memories 12, 22, 32, and 42, respectively, are the first to fourth hosts 10, 20, 30 , 40) may be provided when formed in a ring topology structure through the first to fourth host adapters 100, 200, 300, 400, and the first to fourth hosts 10, 20, 30, 40 It does not limit the type of identification number assigned.

In an embodiment in which the first to fourth hosts 10, 20, 30, and 40 each assign identification numbers to their hosts when a multi-host cluster is formed in a ring topology structure, the first to fourth hosts 10, The first to fourth CPUs (11, 21, 31, 41) of 20, 30, 40) assign identification numbers to their hosts, respectively, and this in the first to fourth memories (12, 22, 32, 42). Each can be saved.

At this time, the first to fourth CPUs (11, 21, 31, 41) form a ring table by matching identification numbers for the first to fourth hosts (10, 20, 30, 40) constituting the ring topology. And, the data is stored in the first to fourth memories 12, 22, 32, and 42, and data can be transmitted/received by referring to the ring table when data is shared between hosts.

In one example, the first CPU 11 generates identification numbers of hosts other than the first host 10 as a ring table and stores them in the first memory 12, and the second CPU 12 is a second host ( 20) The identification numbers of the other hosts are generated as a ring table and stored in the second memory 22, and the third CPU 13 generates the identification numbers of the other hosts other than the third host 30 as a ring table. 3 stored in the memory 32, and the fourth CPU 14 generates identification numbers of the remaining hosts other than the fourth host 40 as a ring table and stores them in the fourth memory 42, and the first to fourth CPUs When (11, 21, 31, 41) shares data between hosts, data transmission/reception between its host and other hosts is performed by referring to the host identification number in the ring table stored in the first to fourth memories (12, 22, 32, 42). Process.

The first to fourth CPUs (11, 21, 31, 41) separately have an identification number for the host to transmit data to another host and an identification number for the host to receive data from another host. In an example, the first CPU 11 may create and refer to a ring table as shown in Table 1 below.

Host Identification number First host sauce 11 destination 12 Second host Send (Host 1 -> Host 2) 21 Incoming (host 1<-host 2) 22 3rd host Send (Host 1 -> Host 3) 31 Incoming (host 1<-host 3) 32 4th host Send (Host 1 -> Host 4) 41 Incoming (host 1<-host 4) 42

As shown in Table 1, the first CPU 11 has an identification number for hosts 20, 30, and 40 other than its own host 10 as a ring table, but the identification number for each host is data with the host. As it is included as an identification number for transmission and reception, the first host 10 transmits data to the third host 30 that is not directly connected through the first host 10 and the first host adapter 100. In this case, data can be shared between the first host 10 and the third host 30 by referring to the identification number 31 of the third host 30 provided in the ring table. Accordingly, in a multi-host cluster formed in a ring topology structure, a PCIe NTB-based interconnect network capable of sharing data between hosts using an identification number is implemented even if the host is not directly connected to the host adapter.

In this case, data sharing between hosts is performed through a method of copying a memory of a data area to be shared or through direct memory access (DMA).

In an embodiment referring to Table 1 and FIGS. 2 and 3 above, when the first host 10 transmits data to the second host 20 directly connected through the host adapter, the source for transmitting data ( source) is the first host 10 and the destination for receiving data is the second host 20, so the first CPU 11 of the first host 10 uses the source identification number (SrcId) as its host It is set to 11, which is an identification number of, and designates a destination identification number (DestId) to receive data as 22, which is a reception identification number of the second host.

At this time, since the first host 10 and the second host 20 have a structure connected to enable direct data transmission, the first CPU 11 transmits the data content to be transmitted to the second host 20. The memory window 120 is written to a memory window 1212 for data transmission among the memory windows 121 for the second host. At this time, the data contents recorded in the memory window 1212 for data transmission are at least some data stored in the first memory 12, and the first CPU 11 is the first memory 12 in which the corresponding data is stored. The memory of the data area of is copied to the memory window 1212.

Alternatively, the first CPU 11 may transmit the data content to be transmitted from the first host 10 to the second host 20 to the second host 20 through DMA. In this case, the first host adapter The data to be transmitted is shared by directly connecting to the second host 20 through (100).

In this way, data is shared through the first memory window 120, which is a shared memory between the first host 10 and the second host 20 connected through the first host adapter 100, or data is shared through DMA. , Information such as source identification number (SrcId), destination identification number (DestId), memory address, and transmission data size are provided through a scratchpad register connected between the first host 10 and the second host 20. It is transmitted from the first host 10 to the second host 20.

In one example, the scratchpad register connecting the first and second hosts 10 and 20 is not shown in the drawing, but is formed between two different hosts in a multi-host cluster formed in a ring topology structure.

In an example in which the first CPU 11 transmits data to the second host 20 through DMA, the first CPU 11 transmits data through DMA, and then through a doorbell register. An interrupt notifying the second host 20 of DMA data transfer is generated, and the data transfer is notified to the other host.

At this time, the host receiving the interrupt can check whether the destination identification number is its own host, and if the data receiver related to the interrupt is its own host, the data directly transmitted to itself from the memory address received through the DMA can be checked.

However, at this time, if the data receiver related to the interrupt is not the host, that is, if the destination identification number (DestId) is different from the identification number of the host, the transmitted data is forwarded to another host, and the destination identification number of the data Data is forwarded until the same host receives it.

As such, as multiple host clusters share data through PCIe NTB-based host adapters, high-speed data sharing becomes possible, and three or more multi-host structures in PCIe NTB-based network structure through ring topology structure. By implementing, there is an effect of providing a high-performance supercomputer at low cost compared to the conventional interconnect technology dependent on technology.

Next, the data sharing method in the PCIe NTB-based multi-host data sharing system having the structure of FIGS. 1 to 3 will be described with reference to FIGS. 4 and 5, the step of transmitting data by the transmitting host (S100) , The memory area copied to the memory window performed according to the step of receiving data by the sending host and the host connected by the host adapter (S200), and the step of determining whether the destination identification number is matched by the host receiving the data (Q100) Checking data through (S310) or forwarding data to a neighboring host (S320).

First, the step of transmitting data by a sending host (S100) is a step in which a specific host transmits data, and will be described in detail with reference to FIG. 5, the source identification number and the receiving host identification number Setting the destination identification number (DestId) (S110), copying the memory area of the data to be transmitted to the memory window of the sending host (S120), information on the source identification number, destination identification number, memory address, and transmission data size Transmitting the data through the scratch pad register (S130), and generating an interrupt informing the transmission host of data transmission (S140).

At this time, the step of setting the sending host identification number and the receiving host identification number as a source identification number (SrcId) and a destination identification number (DestId) (S110) is the identification number of the data transmission host and the data receiving host by the transmission host This is the step of setting by referring to the identification number ring table in Table 1 defined when creating a multi-host cluster.

In the next step (S120), the transmitting host copies the memory area of the data to be transmitted to the memory window of the transmitting host. As the data area is copied to the memory window of the transmitting host, the corresponding data is shared.

And, in the step (S130) of transmitting information on the source identification number, destination identification number, memory address, and transmission data size through the scratch pad register, the information on the data transmitted as the memory area through the memory window of the sending host is scratched. In the step of transmitting through the pad register, and finally, in the step of generating an interrupt notifying data transmission by the sending host (S140), an interrupt notifying data transmission through the doorbell register of the sending host is generated to the connected host through the host adapter. do.

When the sending host transmits data (S100) through these steps (S110, S120, S130, S140), the sending host and the host connected by the host adapter receive the data (S200). At this time, the sending host It is recognized that data has been received by the interrupt generated in S140, and the sending host receives the source identification number and the destination identification number transmitted (S130) through the scratchpad register.

Then, the sending host and the host connected by the host adapter perform a step (Q100) of checking whether the destination to which the sending host intends to transmit data is itself.

At this time, the host receiving the data transmitted by the sending host is a neighboring host connected to the sending host by the host adapter, and the destination to which the sending host intended to send data may or may not be. Is compared with the destination identification number (DestId) (Q100). The host receiving the data compares the destination identification number received through the scratchpad register with its host.

In the above determination step (Q100), if the host that received the data is the same as the destination identification number, Yes, in the direction of the arrow, a step (S310) of checking data through the memory area copied to the memory window is performed. The receiving host (data receiving host) is the final destination to which the host (data transmitting host) that transmitted data attempted to share data, so the data receiving host accesses the memory area copied to the memory window to check the data. Since the memory window to which the data transmission host copied the data area is an address shared with the data reception host, the data reception host can access the memory window shared with the data transmission host to check data.

On the other hand, in the above determination step (Q100), if the host receiving the data is not the same as the destination identification number, a step (S320) of forwarding the data to a neighboring host is performed in the direction of the arrow No. In this case, data reception The host forwards the data area copied to the memory window and the information received through the scratchpad register to the neighboring host. This step (S320) is repeatedly performed until the above step (S310) is performed because the identification number of the data receiving host is the same as the destination identification number of the data transmitting host.

That is, if the data receiving host is not a destination to which the data transmitting host intends to share data, the data receiving host, which has received the data, forwards the shared data until the corresponding shared data reaches the original destination of the data transmitting host.

As described with reference to the flow charts of FIGS. 4 to 5, since data can be shared based on PCIe NTB in a multi-host cluster having a ring topology structure, a high-performance supercomputer is an interconnect technology independent of the prior art. You can expect the effect of sharing data at high speed.

Although the embodiments of the present invention have been described in detail above, the scope of the present invention is not limited thereto, and various modifications and improvements by those skilled in the art using the basic concept of the present invention defined in the following claims are also provided. It belongs to the scope of rights.

1, 1a: PCIe NTB-based data sharing system between multiple hosts
10: first host 11: first CPU
12: first memory 100: first host adapter
101: 1-1 NTB port 102: 1-2 NTB port
120: first memory window 121: memory window for the second host
122: memory window for the third host
1211: Memory window for data reception
1212: Memory window for data transmission

Claims (11)

An independent computer that transmits or receives data, comprising at least three hosts each including a CPU that processes data transmission or reception, and a memory that stores data or allocates a memory window that is a separate address area; and
Each connected to the at least three hosts, each having at least two NTB ports, one of the at least two NTB ports is connected to an NTB port of a host adapter connected to the neighboring host, and the other is another of the At least three PCIe NTB-based host adapters, each connected to the NTB port of the host adapter connected to the neighboring host, and connected to two different neighboring hosts in a ring topology structure through the NTB port
Including,
PCIE NTB-based multi-host data sharing system, characterized in that the memory window is shared through NTB ports of host adapters formed in the two hosts to transmit and receive data between the two hosts.
The method of claim 1,
The memory windows are respectively formed for two different neighboring hosts connected through a host adapter of a host on which the memory window is formed,
PCIE NTB-based data sharing system between multiple hosts, characterized in that data is shared between two hosts through memory windows respectively formed for the two different neighboring hosts.
The method of claim 1,
The memory windows each include a memory window for receiving data and a memory window for transmitting data,
PCIE NTB-based data sharing system between multiple hosts, characterized in that data is shared between two hosts through the memory window according to data transmission or reception by a host.
The method of claim 2,
The memory windows respectively formed for the two different neighboring hosts each include a memory window for receiving data and a memory window for transmitting data,
PCIE NTB-based data sharing system between multiple hosts, characterized in that data is shared between two hosts through the memory window according to data transmission or reception by a host.
The method of claim 1,
PCIE NTB-based data sharing system between multiple hosts, characterized in that the data transmission and reception between the two hosts is performed through DMA.
The method of claim 5,
PCIE, characterized in that the host transmitting data generates an interrupt notifying the DMA data transmission through the doorbell register after transmitting the data through the DMA, and the host receiving the data checks the data received through the interrupt. NTB-based data sharing system between multiple hosts.
The method of claim 1,
The CPU forms the identification numbers of the at least three hosts as a ring table, and when the host transmits data to another host, the identification number of the host transmitting data and the host to receive the data refer to the ring table. PCIE NTB-based data sharing system between multiple hosts, characterized in that transmitting the identification number as information.
The method of claim 7,
The information further includes a memory address of data to be shared and a size of transmission data. PCIE NTB-based data sharing system between multiple hosts, characterized in that.
The method of claim 7,
Further comprising a scratch pad register connecting between the host transmitting the data and the host receiving the data,
The host transmitting the data transmits the information to the host receiving the data through the scratchpad register.
The method of claim 7,
In a ring topology between host adapters in which the at least three hosts are not directly connected through the host adapter,
As a specific host receives the data and the information, the identification number of the host that will receive the data included in the information is compared with the identification number of its own host, and if it matches, the data is read from the memory window and matched. If not, a PCIE NTB-based data sharing system between multiple hosts, characterized in that the received data is forwarded to a neighboring host.
In a method for sharing data in a data sharing system between multiple hosts based on PCIE NTB comprising at least three hosts and at least three host adapters each connected to the at least three hosts and including at least two NTB ports,
A step in which the sending host transmits data through the host adapters, and transmits the identification number of the sending host and the destination host identification number of the data as information (S100),
Receiving data by a host connected to the sending host (S200),
A step (Q100) of determining, by the host receiving the data, whether the destination host identification number included in the information matches its host identification number, and
As a result of performing the determination step (Q100), when the destination host identification number included in the information matches its host identification number, the host receiving the data checks the data through the memory area copied to the memory window (S310) and, as a result of performing the determination step (Q100), as the destination host identification number included in the information is inconsistent with its own host identification number, the host receiving the data forwards the data to the neighboring host ( Step of performing any one of S320)
PCIE NTB-based data sharing method between multiple hosts, comprising a.

Images