CN201018634Y - Tree-type optical packet switching cluster structure - Google Patents

Abstract

The utility model discloses a tree-type optical packet switching cluster structure, which comprises at least two groups of CPUs (A), and an optical interconnection switching device is connected between the CPUs (A), and the optical interconnection switching device includes at least the first level For nodes (1) and second-level nodes (2), the current HPCS mostly adopts a centralized control method. After each port has a service, it must send a request signal to the central scheduler. If the distance is far away, it may introduce a request and reply. The round-trip delay of the signal is large. The utility model adopts a hierarchical control management mode, and the nodes only need to send requests to the local control unit, which can greatly reduce such time delay, and at the same time, can reduce the storage and calculation pressure of each level of control unit.

Description

Tree-type optical packet switching cluster structure

technical field

The utility model relates to a hierarchical control computer, in particular to a tree-type optical packet switching cluster structure.

Background technique

In the past few years, in the field of technology, the development of computing resources often cannot keep up with the needs of applications. In some specialized fields such as quantum physics experiments and global climate simulations, it is necessary to share a large amount of measurement data to strengthen the ability of analysis and calculation. At the same time, in order to meet the needs of industries and fields such as scientific research, resource environment, manufacturing and service industries, it is hoped to develop a number of large-scale application systems based on high-performance computing and grid technology to realize resource sharing and collaboration in industries and fields Work, improve the productivity of the industry, and promote the informatization construction of various departments.

Traditional computing and communication resources are no longer sufficient. During the rapid development of high-performance computing, people have proposed the idea of connecting different computing resources in different regions through high-speed networks to solve some "mass-level" applications. The "grid" concept proposed in the mid-1990s is the main way to realize this idea.

However, the bandwidth of electricity is affected by factors such as its inherent high-frequency capacitance, inductance effect, and electromagnetic interference, which has become a bottleneck for high-speed, high-capacity, and low-latency data transmission.

Now, optical fiber and DWDM technologies are maturing day by day, and distributed computing has been brought to a new level. Through high-speed optical network and optical interconnection technology, the idea of integrating a series of distributed computing and storage nodes as a large computing and storage unit is technically and economically feasible.

At present, the design of HPCS and Grid can be divided into Optical Circuit Switching (OCS), Optical Burst Switching (OBS) and Optical Packet Switching (OPS) from the perspective of optical switching.

OCS-based HPCS attempts to establish a permanent static optical path connection between nodes, which is suitable for high-throughput traffic transmission, but its static characteristics reduce the utilization of ports and links, and cannot meet the needs of users for burst services .

The use of OBS technology can not only meet the requirements of higher throughput for the transmission of larger files, but also meet the requirements of lower latency for the transmission of smaller commands. However, the burst assembly will introduce a large burst assembly delay, which cannot meet the low-latency real-time transmission requirements.

Contents of the invention

The utility model provides a tree-type optical packet switching cluster structure capable of reducing time delay, satisfying user business sudden demands, and improving link and port utilization.

The utility model adopts the following technical solutions:

Taking two-level HPCS as an example, the first level is defined as the master node, and the second level is the slave node. The master node has n slave nodes; each slave node has m CPUs. The master node has the highest management level, manages and controls n slave nodes, and the slave nodes manage their own m CPUs, and so on, which can achieve multi-level expansion. Define the signal forwarded from the slave node to the master node as an uplink signal, and the signal forwarded from the master node to the slave node as a downlink signal.

The signals generated by the m CPUs of each slave node are connected to the switching unit of the slave node through Dense Wavelength Division Multiplexing (DWDM) technology, and similarly, the signals generated by n slave nodes are connected to the master node through DWDM technology.

Through the control channel, the CPUs regularly transmit information such as their free storage space and free computing power to the slave nodes, and the slave nodes also regularly report the information of the CPUs they manage to the master node. The master and slave nodes store these information in the storage matrix of their own control unit.

After the CPUs have a task, they are assembled into an optical packet at the edge router (ER). The packet header contains information such as the storage capacity and computing power required by the task. If there is a clear destination, it contains address information.

The packets are first exchanged inside the slave node where the CPU is located. The control unit allocates or finds the destination for the task according to the packet header information. If there is no destination unit that meets the requirements, the packet is forwarded to the master node.

The master node balances according to the situation of each slave node, finds a suitable slave node, and forwards the task to the corresponding slave node, and the slave node that obtains the task allocates the appropriate destination CPUs for the task according to the packet header information.

The optical switching unit adopts multicast technology to realize high-speed and large-capacity data exchange.

Compared with the prior art, the utility model has the following advantages:

The utility model can realize the interconnection of large batches of distributed high-performance microprocessors and the real-time interaction and sharing of a large number of burst high-speed optical data streams. High-speed and large-capacity data is stored and processed in a distributed control and hierarchical management mode, which can reduce the calculation and storage pressure of the control unit, facilitate the realization of multi-level expansion, and facilitate the connection between long-distance CPUs. The use of asynchronous switching mode, OPS and multicast technology can reduce the delay, meet the sudden demand of user services, improve the utilization rate of links and ports and the efficiency of data transmission, while avoiding clock synchronization.

The current HPCS mostly adopts a centralized control method. After each port generates a service, it must send a request signal to the central scheduler. If the distance is long, a large round-trip delay for the request and reply signals may be introduced. The utility model adopts the hierarchical control management mode, the node only needs to send a request to the local control unit, which can greatly reduce such time delay, and at the same time, can reduce the storage and calculation pressure of each level of control unit;

The utility model adopts the asynchronous exchange mode, which can reduce the data waiting time, meet the sudden demand of user business, avoid the demand for synchronous clock signal, and increase the flexibility of business exchange. Especially when performing multi-level expansion, if synchronous exchange is used, the synchronization between nodes at all levels is more difficult, and the asynchronous exchange mode can avoid this problem;

The utility model adopts DWDM technology, which can make full use of the characteristics of large capacity of optical fiber, avoid the transmission bottleneck of small electrical signal capacity, and improve the utilization rate of the channel. In addition, the optical fiber has better anti-interference ability and price compared with the electric transmission and exchange technology. Incomparable advantages;

The utility model adopts OPS technology to realize data exchange, which can not only avoid the lower port and link utilization rate caused by the static characteristics of OCS, but also avoid the large assembly time delay introduced by burst assembly in OBS, and can meet the needs of user services. Bursty and real-time transmission requirements;

Compared with point-to-point unicast technology, the utility model adopts multicast technology, which can greatly improve data transmission efficiency;

When the current HPCS structure implements multi-level expansion, each level often needs to perform photoelectric-optical conversion, and the switching speed and throughput will be affected. When the multi-level control technology adopted by the utility model is used for multi-level expansion, multiple photoelectric-optical conversions are not required, and the throughput can be greatly improved and the time delay can be reduced.

Description of drawings

Fig. 1 is a schematic diagram of the structure of the two-stage system of the utility model for the on and off roads.

Fig. 2 is a structural schematic diagram of the node control unit of the present invention.

Fig. 3 is a schematic structural diagram of an m-level system according to an embodiment of the present invention.

In the figure, A, CPUs, 11, first-level node control unit, 12, first-level node optical switching unit, 13, first-level node optical multiplexer, 14, first-level node optical beam splitter, 15, first-level node optical beam splitter, First-level node fiber delay line, 211, second-level first node optical multiplexer, 212, second-level first-node optical multiplexer, 213, second-level first-node optical beam splitter, 214, The optical switching unit of the first node of the second level, 215, the control unit of the first node of the second level, 216, the optical switching unit of the first node of the second level, 217, the edge router, 218, the first node of the second level Optical fiber delay line, A11, wavelength conflict resolution module for the drop channel of the first node in the second stage, A12, wavelength conflict resolution module for the add channel of the first node in the second stage, An1, wavelength conflict resolution module for the drop channel of the nth node in the second stage Module, An2, the on-line wavelength conflict resolution module of the nth node in the second level.

Detailed ways

Example 1

The utility model proposes a novel hierarchical control HPCS based on OPS and optical multicast. Taking two-level HPCS as an example, the first level is defined as the master node, and the second level is the slave node. The master node has n slave nodes; each slave node has m CPUs. The master node has the highest management level, manages and controls n slave nodes, and the slave nodes manage their own m CPUs, and so on, which can achieve multi-level expansion. Define the signal forwarded from the slave node to the master node as an uplink signal, and the signal forwarded from the master node to the slave node as a downlink signal.

The signals generated by the m CPUs of each slave node are connected to the switching unit of the slave node through Dense Wavelength Division Multiplexing (DWDM) technology, and similarly, the signals generated by n slave nodes are connected to the master node through DWDM technology.

Through the control channel, the CPUs regularly transmit information such as their free storage space and free computing power to the slave nodes, and the slave nodes also regularly report the information of the CPUs they manage to the master node. The master and slave nodes store these information in the storage matrix of their own control unit.

After the CPUs have a task, they are assembled into an optical packet at the edge router (ER). The packet header contains information such as the storage capacity and computing power required by the task. If there is a clear destination, it contains address information.

The packets are first exchanged inside the slave node where the CPU is located. The control unit allocates or finds the destination for the task according to the packet header information. If there is no destination unit that meets the needs, the packet is forwarded to the master node through the optical add/drop unit.

The optical add/drop unit consists of a tunable fiber Bragg grating and an optical circulator.

The master node balances according to the situation of each slave node, finds a suitable slave node, and forwards the task to the corresponding slave node, and the slave node that obtains the task allocates the appropriate destination CPUs for the task according to the packet header information.

The optical switching unit adopts multicast technology to realize high-speed and large-capacity data exchange.

With reference to Fig. 1, a kind of tree-type optical packet switching cluster structure, comprises at least 2 groups of CPUA, is connected with optical interconnection switching device between CPUA, optical interconnection switching device comprises at least the first level node 1 and the second level node 2, of which:

The first-level node is composed of the first-level control unit 1 ₁ , the first-level optical switching unit 1 ₂ , the first-level optical multiplexer 1 ₃ and the first-level optical beam splitter 1 ₄ , the first-level optical multiplexer 1 ₃ The output end of the first-stage optical beam splitter ₁₄ is connected to the input end of the first-stage optical beam splitter 14, and one output end of the first-stage optical beam splitter ₁₄ is connected to the input end of the first-stage control unit ₁₁ , which is used to connect the The signal sent to the first-level control unit 1 ₁ , the other output end of the first-level optical beam splitter 1 ₄ is connected to the input end of the first-level optical switching unit 1 ₂ through an optical fiber delay line 1 ₅ for the The signal containing data is transmitted to the first-level optical switching unit, and the control signal output end of the first-level control unit ₁₁ is connected to the control signal input end of the first-level optical switching unit ₁₂ for controlling the first-level optical switching unit 1 ₂ for data multicast exchange;

The second-level node includes at least n second-level nodes, n≥2 and n is a natural number, and the second-level node 2 includes a second-level optical multiplexer 21 ₁ , a second-level multiplexer 21 ₂ , a second-level optical beam splitter 21 ₃ , optical add/drop unit 21 ₄ , second-level node control unit 21 ₅ , and second-level node optical switch unit 21 ₆ , the bus outputs of the CPUs respectively pass through the signal sending end of the edge router 21 ₇ and the second-level optical complex The input end of the multi-channel optical multiplexer 21 ₁ is connected to the optical fiber using Dense Wavelength Division Multiplexing (DWDM) technology, and the output end of the second-stage optical multiplexer 21 ₁ is connected to the optical fiber and the multiplexer 21 ₂ connected to one input end of the first-level optical switching unit 12, and the output ends of the above-mentioned first-level optical switching unit ₁₂ pass through the drop wavelength conflict resolution modules _A11 , ..., An ₁ and the other multiplexers ₂₁₂ , ... 2n ₂ of each second-level node respectively. One input end is connected, and the multiplexer ₂₁₂ is used to multiplex the drop signal generated by the first-level optical switching unit ₁₂ into the second-level node, and the output end of the multiplexer ₂₁₂ is connected to the second-level optical beam splitter The input end of ₂₁₃ is connected, and an output end of the second-level optical beam splitter ₂₁₃ is connected with an input end of the second-level optical add/drop unit ₂₁₄ through an optical fiber delay line ₂₁₈ , and the second-level optical beam splitter The other output end of ₂₁₃ is connected to the control signal input end of the second-level node control unit ₂₁₅ , and the second-level optical beam splitter ₂₁₃ is used to separate the data signal from the packet header control signal, and the optical add/drop unit ₂₁₄ The output end is connected to the input end of the second-level node optical switching unit ₂₁₆ , and the output ends of the second-level node optical switching unit ₂₁₆ are respectively connected to the signal receiving end of the second-level edge router, and the received optical signal is converted into The electrical signal is transmitted to the memory of the CPUs, and a control signal output end of the second-level node control unit ₂₁₅ is connected to the control signal input end of the optical add/drop unit ₂₁₄ , which is used to control the state of the optical add/drop unit ₂₁₄ . The data exchanged at the first level and the second level are separated, and another control signal output port is connected with the control signal input end of the second level node optical switching unit ₂₁₆ for controlling the second level optical switching unit ₂₁₆ to perform In the multicast exchange of data, the drop signal output ends of each second-level optical add/drop unit ₂₁₄ are respectively connected to the input ends of the first-level optical multiplexer ₁₃ through the add wavelength conflict resolution modules Al ₂ , ..., An ₂ , It is used to upload the data that needs to _be exchanged at _the first-level node to the first-level node. Referring to FIG. The control unit is composed of a photoelectric converter B1, a head recognition unit B2, a control unit interface B3 and a storage matrix B4. The input terminal of the photoelectric converter B1 is used as the control signal input terminal of the control unit, and the output terminal of the photoelectric converter B1 is Connect to the input end of the head identification unit B2, the output end of the head identification unit B2 is respectively connected to the input end of the control unit interface B3 and the storage matrix B4, and the head identification unit B2 directly transmits the control signal with a clear destination address to the control unit interface B3, and the control signal without the destination address is transmitted to the storage matrix B4, and the polling to the storage matrix B4 is the task assignment destination, the output end of the storage matrix B4 is connected with the control unit interface B3 input end, and the control unit interface B3 The output end serves as the control signal output end of the control unit.

Example 2

Referring to Figure 3,

An m-level computer system based on optical packet switching and optical multicasting, comprising n ^(m-1) groups of CPUs, connected with optical interconnection switching devices between the CPUs, characterized in that it comprises m-level optical interconnection switching devices, in:

The first-level node is composed of the first-level control unit, the first-level optical switching unit, the first-level optical multiplexer and the first-level optical beam splitter. The output of the first-level optical multiplexer and the first-level optical beam splitter connected to the input end of the first-level optical beam splitter, and an output end of the first-level optical beam splitter is connected to the input end of the first-level control unit for transmitting the signal containing the packet header to the first-level control unit, and the first-level optical beam splitter The other output end of the device is connected to the input end of the first-level optical switching unit through an optical fiber delay line, and is used to transmit the signal containing data to the first-level optical switching unit, and the control signal output end of the first-level control unit is connected to the first-level optical switching unit. The control signal input end of the first-level optical switching unit is connected to control the first-level optical switching unit to perform data multicast switching;

The second-level nodes include at least n second-level nodes, n≥2 and n is a natural number, and the second-level nodes include the second-level optical multiplexer, the second-level combiner, the second-level optical beam splitter, and optical add/drop unit, the second-level node control unit and the second-level node optical switching unit, and the output ends of the drop signal of the optical add/drop unit of the third-level node are respectively connected to the input ends of the second-level optical multiplexer through the add wavelength conflict resolution unit , used to multiplex the add-on signal of the third-level node into the second-level node using Dense Wavelength Division Multiplexing (DWDM) technology, and the output end of the second-level optical multiplexer is connected to an input end of the multiplexer through an optical fiber, The output ends of the above-mentioned first-level optical switching units are respectively connected to the other input ends of the multiplexers of each secondary node through the drop wavelength conflict resolution module, and the multiplexers are used to combine the drop wavelength generated by the first-level optical switching unit. The signal is multiplexed into the second-level node, the output end of the multiplexer is connected to the input end of the second-level optical beam splitter, and one output end of the second-level optical beam splitter is connected to the second-level light through the fiber delay line One input end of the road unit is connected, and the other output end of the second-level optical beam splitter is connected with the control signal input end of the second-level node control unit. The optical beam splitter is used to separate the data signal from the packet header control signal. The output end of the add/drop unit is connected to the input end of the second-level node optical switching unit, and the output end of the second-level node optical switching unit is respectively connected to the signal receiving end of the second-level edge router, and the received optical signal is converted into The electrical signal is sent to the memory of the CPUs, and a control signal output terminal of the second-level node control unit is connected to the control signal input terminal of the optical add-on and drop-off unit to control the state of the optical add-on and add-off unit. The data exchanged by the second stage is separated, and the other control signal output port is connected with the control signal input port of the second-stage node optical switching unit, which is used to control the second-stage optical switching unit to perform data multicast exchange, and each second-stage The output ends of the drop signal of the optical add/drop unit are respectively connected to the input ends of the first-level optical multiplexer through the add wavelength conflict resolution module, and are used to add the data that needs to be exchanged at the first-level node to the first-level node.

The m-level node has the same structure as the second-level node, including n ^(m-1) m-level nodes, n, m≥2 and n and m are natural numbers, and the m-level node includes the m-level optical multiplexer, the m-level multiplexer, m-level optical beam splitter, m-level optical add/drop unit, m-level node control unit and m-level node optical switching unit, the bus output ends of CPUs respectively pass through the signal sending end of the edge router It is connected to the input end of the mth-level optical multiplexer, which is used to multiplex multiple optical signals into the optical fiber using Dense Wavelength Division Multiplexing (DWDM) technology, and the output end of the m-th One input end is connected, and the output end of the m-1th level optical switching unit is respectively connected to the other input end of the multiplexer of each m-level node through the drop wavelength conflict resolution module, and the m-level multiplexer is used to combine the mth -The drop signal generated by the first-level optical switching unit is multiplexed into the m-th-level node, the output of the m-th-level multiplexer is connected to the input of the m-th-level optical beam splitter, and one of the m-level optical beam splitters The output end is connected to one input end of the mth-level optical add/drop unit through the fiber delay line, the other output end of the m-level optical beam splitter is connected to the control signal input end of the m-level node control unit, and the m-level optical The beam splitter is used to separate the data signal from the packet header control signal. The output end of the mth level optical add/drop unit is connected to the input end of the mth level node optical switching unit, and the output end of the mth level node optical switching unit is connected to the output end of the mth level node optical switching unit respectively. The signal receiving end of the m-level edge router is connected, and the received optical signal is converted into an electrical signal and sent to the memory of the CPUs. A control signal output terminal of the m-level node control unit is connected to a control signal input of the m-level optical add/drop unit. terminal connection, used to control the status of the m-th level optical add/drop unit, separate the data that needs to be exchanged at the m-1 level and the m-th level respectively, and the other control signal output port is connected with the control of the m-th level node optical switching unit The signal input terminal is connected to control the m-th level optical switching unit for multicast data exchange. The drop signal output terminals of each m-th level optical add/drop unit respectively pass through the add wavelength conflict resolution module and the m-1 level optical complex connected to the input end of the user, and is used to upload the data to be exchanged by the m-1th level node to the m-1th level node.

Claims (2)

1. a tree type light packet switching formula group of planes structure comprises at least 2 group CPU (A), is connected with light interconnection switch between CPU (A), it is characterized in that light interconnection switch comprises the 1st grade of node (1) and the 2nd grade of node (2) at least, wherein:

The 1st grade of node is by first order control unit (1 ₁), first order light crosspoint (1 ₂), first order optical multiplexer (1 ₃) and first order beam splitter (1 ₄) form first order optical multiplexer (1 ₃) output and first order beam splitter (1 ₄) input connect first order beam splitter (1 ₄) output and first order control unit (1 ₁) input connect, the signal that is used for containing packets headers is sent to first order control unit (1 ₁), first order beam splitter (1 ₄) another output by fiber delay line (1 ₅) and first order light crosspoint (1 ₂) input connect, the signal that is used for containing data is sent to first order light crosspoint, first order control unit (1 ₁) control signal output ends and first order light crosspoint (1 ₂) signal input end connect, be used to control first order light crosspoint (1 ₂) to carry out the multicast exchange of data;

The 2nd grade of node comprises n two-level node at least, and n 〉=2 and n are natural number, and this two-level node (2) comprises second level optical multiplexer (21 ₁), second level wave multiplexer (21 ₂), second level beam splitter (21 ₃), light top and bottom path unit (21 ₄), node control unit, the second level (21 ₅) and second level node optical crosspoint (21 ₆), the output end of main of CPUs passes through edge router (21 respectively ₇) signal sending end and second level optical multiplexer (21 ₁) input connect, be used for multipath light signal adopted that dense wave division multipurpose (DWDM) technology is multiplexing advances optical fiber, second level optical multiplexer (21 ₁) output through optical fiber and wave multiplexer (21 ₂) input connect above-mentioned first order light crosspoint (1 ₂) output respectively through following road wavelength conflict-solving module (Al ₁..., An ₁) with the wave multiplexer (21 of each two-level node ₂... 2n ₂) another input connect this wave multiplexer (21 ₂) be used for first order light crosspoint (1 ₂) the following road signal multiplexing that produces advances two-level node, wave multiplexer (21 ₂) output and second level beam splitter (21 ₃) input connect second level beam splitter (21 ₃) output through fiber delay line (21 ₈) and light top and bottom path unit, the second level (21 ₄) input connect second level beam splitter (21 ₃) another node control unit, output second level (21 ₅) signal input end connect second level beam splitter (21 ₃) be used for data-signal being separated light top and bottom path unit (21 with the packets headers control signal ₄) output and second level node optical crosspoint (21 ₆) input connect second level node optical crosspoint (21 ₆) output be connected with the signal receiving end of second level edge router respectively, and the light signal that receives is converted to the internal memory that the signal of telecommunication is sent to CPUs, node control unit, the second level (21 ₅) control signal output ends and light top and bottom path unit (21 ₄) signal input end connect, be used to control light top and bottom path unit (21 ₄) state, with needs respectively separately in the data of the first order and second level exchange, another control signal output ends and second level node optical crosspoint (21 ₆) signal input end connect, be used to control second level light crosspoint (21 ₆) to carry out the multicast exchange of data, each light top and bottom path unit, second level (21 ₄) following road signal output part respectively through the wavelength conflict-solving module (Al that sets out on a journey ₂..., An ₂) and first order optical multiplexer (1 ₃) input connect, be used for and need set out on a journey to first order node in the data of first order node switching.

2. tree type light packet switching formula group of planes structure according to claim 1 is characterized in that first order node control unit (1 ₁), two-level node control unit (21 ₅), or m level node control unit adopts control unit to realize, this control unit is by optical-electrical converter (B1), recognition unit (B2), control unit interface (B3) and storage matrix (B4) are formed, the input of optical-electrical converter (B1) is as the signal input end of control unit, the output of optical-electrical converter (B1) is connected with the input of a recognition unit (B2), the output of recognition unit (B2) is connected with the input of control unit interface (B3) and storage matrix (B4) respectively, the control signal that recognition unit (B2) will have the address that has a definite purpose directly transfers to control unit interface (B3), and the control signal that will not have destination address transfers to storage matrix (B4), by storage matrix (B4) is polled as the Task Distribution destination, the output of storage matrix (B4) is connected with control unit interface (B3) input, and the output of control unit interface (B3) is as the control signal output ends of control unit.

Images