CN101505319B - Method for accelerating adaptive reconfigurable processing unit array system based on network - Google Patents

Abstract

The invention discloses a method for speeding up an adaptive reconfigurable processing unit array system based on a network. It is a computing node composed of general-purpose processors and reconfigurable devices to form an adaptive reconfigurable processing unit network. The program to be executed is divided into a set of tasks that can run independently. Each task in the task set is distributed through an adaptive reconfigurable processing unit network to find suitable computing resources for execution, and utilize the high efficiency and Flexible features, reconfiguration, so that the functions of the reconfigurable device can meet the needs of different tasks. Under the action of the adaptive reconfigurable processing unit network, the network-based adaptive reconfigurable processing unit array system will intelligently allocate tasks according to the conditions of the nodes on the network, and give priority to assigning tasks to other computing resources that are idle. Node, thereby reducing the execution time of the program, improving the utilization rate of the reconfigurable device, and achieving the purpose of program acceleration.

Description

Method for Acceleration of Network-Based Adaptive Reconfigurable Processing Unit Array System

technical field

The invention relates to the field of multi-core technology and the field of reconfigurable technology, in particular to a method for implementing a reconfigurable multi-core accelerated program execution based on network communication.

Background technique

With the continuous improvement of computer manufacturing technology, VLSI technology has made great progress, and multi-core technology has become the current mainstream processor technology. However, as the number of processor cores continues to increase, how to improve the utilization of these cores has become a difficult problem.

Since the parallelism of general application programs is not high, it is generally believed that after the number of general-purpose cores of the processor exceeds 16, it is difficult to improve performance by increasing the number of general-purpose processing cores.

In heterogeneous multi-core processors, some special-purpose cores specially customized for special tasks can be integrated, so as to achieve the high performance of customized applications, but this kind of customized heterogeneous multi-core processors can only achieve good performance when running the targeted applications Performance, otherwise the performance is not as good as the general processing core.

The emergence of reconfigurable technology makes general-purpose processors integrate FPGA-based reconfigurable cores to provide another solution. While retaining the high flexibility of general-purpose processors to adapt to most applications, the high performance and high efficiency of special-purpose processors are obtained. At the same time, the reconfigurable core can support various types of applications through reconfiguration, so that it can be more flexibly applied to various programs.

But an isolated node does not have enough power and flexibility to handle some applications.

Insufficient reconfigurable resources: When the application needs more reconfigurable resources than an isolated node has, in this case, some requests of the application must be suspended until the occupied reconfigurable resources are It must be freed and reconfigured before proceeding, which reduces performance and efficiency.

Waste of reconfigurable resources: When the application program is not suitable for reconfigurable devices, the reconfigurable resources of the node are idle and wasted.

High cost of frequent reconfiguration: When the functional modules frequently called by the application cannot be provided on the reconfigurable device at the same time, this requires frequent switching of the functional modules on the reconfigurable device, resulting in frequent reconfiguration, which makes Performance and efficiency drop.

Contents of the invention

The purpose of the present invention is to provide a network-based acceleration method for an adaptive reconfigurable processing unit array.

The technical scheme that the present invention solves its technical problem adopts is as follows:

1) Task division of the original program:

The system divides a program into a collection of tasks that can run independently;

2) Build an adaptive reconfigurable processing unit array based on the macro network:

The network-based adaptive reconfigurable processing unit array is composed of n nodes, n ∈ [1, 2, 3...), each node has 0-4 neighbor nodes directly connected via Ethernet, and each node is composed of It consists of two parts, the first part is a general-purpose computer, and the second part is a reconfigurable device NetFPGA;

Initialize the reconfigurable device according to the tasks to be performed on the reconfigurable device;

3) Distribution of tasks:

For any node in the macro-network-based adaptive reconfigurable processing unit array, itself is a local node, and other nodes are remote nodes;

Task distribution is to distribute the tasks in the task set in the program to the computing resources of the local node and the computing resources of the remote node;

4) Changes in the function of reconfigurable equipment:

For the reconfigurable device of any one of the n nodes, if the resource of the reconfigurable device is idle, and the logic function module required by the assigned task is not currently available, the reconfigurable device needs to be reconfigured functional module;

5) Execution of tasks:

Execute the task after the task is assigned, and return the execution result after the task execution ends.

The steps of building a network-based adaptive reconfigurable processing unit array are as follows:

1) The adaptive reconfigurable processing unit array built is based on a high-speed local area network with a transmission rate of 100Mb or 1000Mb;

2) The self-adaptive reconfigurable processing unit array is composed of n nodes, n∈[1, 2, 3...), each node is composed of two parts, the first part is a general-purpose computer, and the second Some are reconfigurable devices;

3) The reconfigurable device NetFPGA has 4 Ethernet interfaces, so the reconfigurable device of one node can be directly connected with up to four other nodes through a high-speed local area network for communication. The directly connected nodes are called neighbor nodes. Can communicate through neighbor nodes and non-neighbor nodes;

4) The two parts of a node, the general-purpose computer and the reconfigurable device, are connected through a standard PCI interface for interconnection of peripheral components;

5) Customize the control module SuperBlock on the reconfigurable device;

6) Initialize the functional modules on the reconfigurable device.

The steps for changing the function of the reconfigurable device are as follows:

1) Before the entire network-based adaptive reconfigurable processing unit array system starts to work, according to the tasks to be run on the reconfigurable device, generate configuration files for the reconfigurable device;

2) After the system starts working, with the dynamic distribution of tasks, when the logical function block required by the task to be run by the reconfigurable device of any one of the n nodes is not available on the reconfigurable device, the reconfigurable The configuration device sends a reconfiguration request to the host of its own node;

3) After the host receives the reconfiguration request sent by the reconfigurable device of its own node, the host invokes a program to reconfigure the reconfigurable device of its own node.

The control module SuperBlock on the customized reconfigurable device serves as the function of the reconfigurable device inside any of the n nodes and the communication controller of the general computer; Neighboring nodes can reconfigure the functions of the communication controller between devices; manage the assignment of tasks received by the reconfigurable devices during system operation, process the tasks, and record the status of the tasks.

Compared with the background technology, the present invention has the beneficial effects that:

The present invention is a reconfigurable multi-core accelerated program execution method based on network communication, and its main function is to use reconfigurable devices to enhance the computing power of general processors and accelerate the execution of specific function modules. Using network communication to find idle computing resources on the network, on the one hand, it can relieve the pressure on nodes with insufficient computing resources, and on the other hand, it can improve the resource utilization of all nodes on the entire network.

(1) Adaptability: The system will intelligently allocate tasks according to the conditions of nodes on the network. Tasks on busy nodes waiting for resources are assigned to idle nodes.

(2) Efficiency: Since the system will perform task intelligence according to the situation of the nodes on the network, it can utilize the idle resources in the network, thereby improving the performance and efficiency of the application program. The experiment proves that the system can effectively improve the performance.

Description of drawings

Fig. 1 is a system configuration diagram of an example of the present invention.

Figure 2 is a node in the system, which is composed of a general-purpose computer and one or more reconfigurable devices, and the two components are connected by PCI.

Fig. 3 is a flowchart showing the flow of task distribution according to the present invention.

Detailed ways

The method involves related symbol interpretation:

C1: able to run on reconfigurable devices;

C2: The general processor of the local node is idle;

C3: The reconfigurable device of the local node is idle;

C4: There is a neighbor node reconfigurable device idle;

The specific implementation flow of the network-based adaptive reconfigurable processing unit array acceleration system is as follows.

Step 1: Taskization of the original program:

1: The system divides a program into two sets {{A}, {B}} of tasks that can run independently. Set A represents tasks that cannot be run on reconfigurable devices, and set B represents tasks that cannot be run on reconfigurable devices. Tasks running on the device.

2: Generate a dependency table between tasks according to the dependencies between tasks.

Step 2: Build a network-based adaptive reconfigurable processing unit array:

As shown in Figure 1, the adaptive reconfigurable processing unit array built is based on a high-speed local area network with a transmission rate of 100Mb or 1000Mb.

The self-adaptive reconfigurable processing unit array is composed of n nodes, n∈[1, 2, 3…), each node is composed of two parts, the first part is a general-purpose computer, and the second part is Reconfigurable devices.

As shown in Figure 2, the two parts of a node, the general-purpose computer and the reconfigurable device, are connected through a standard PCI interface for interconnecting peripheral components.

The reconfigurable device NetFPGA has 4 Ethernet interfaces. One node can be directly connected with up to four other nodes through a high-speed local area network for communication. The directly connected nodes are called neighbor nodes, and nodes can communicate through neighbor nodes and non-neighbor nodes to communicate.

Customize the control module SuperBlock on the reconfigurable device, that is, write the logic function module. This logic function module is responsible for the communication between the reconfigurable device inside the node and the general computer, and the reconfigurable device of the node and the reconfigurable device of the neighbor node. In the process of system operation, the task distribution received by the reconfigurable device is managed, the task is processed, and the status of the task is recorded. The control module SuperBlock is fixed on the reconfigurable device. When the reconfigurable device is reconfigured, the part of the control module on the reconfigurable device remains unchanged.

Initialize the functional modules on the reconfigurable device, that is, predict the tasks to be performed on the reconfigurable device, and burn the binary stream file of the corresponding functional module to the reconfigurable device, so as to achieve the purpose of initialization.

Step 3: Distribution of tasks:

For any node in the network-based adaptive reconfigurable processing unit array, it is a local node, and other nodes are remote nodes.

The distribution of tasks is to distribute the tasks in the task set in the program to the computing resources of the local node and the computing resources of the remote nodes. Whenever a task is assigned, the task is removed from the collection to which the task belongs. For any node in the network-based adaptive reconfigurable processing unit array, there are two kinds of tasks to be processed, the first is from the local node task collection, and the second is the task forwarded by neighbor nodes. Figure 3 shows the process of task distribution for these two tasks.

When any task in the local node task set distributes tasks:

1: The task to be distributed satisfies the condition C1, and this task enters the waiting queue 2.

2: The task to be distributed does not meet the condition C1, and this task enters the waiting queue 1.

3: When the condition C2 is satisfied, a task taken from the waiting queue 1 is executed.

4: When condition C2 is not satisfied, a task taken from waiting queue 1 re-enters waiting queue 1.

When a task forwarded by a neighbor node distributes tasks:

1: The task for distribution enters the waiting queue 2.

2: When the condition C3 is satisfied, a task taken from the waiting queue 2 is executed.

3: When the condition C3 is not satisfied and the condition C4 is satisfied, a task taken from the waiting queue 2 is forwarded to the neighbor.

4: When the condition C3 is not satisfied and the condition C4 is not satisfied, a task taken from the waiting queue 2 re-enters the waiting queue 2.

Step 4: Changes in the function of the reconfigurable device:

1: Before the entire network-based adaptive reconfigurable processing unit array system starts working, according to the task to be run on the reconfigurable device, write the logic function module corresponding to the task. Since the number of hardware resources occupied by each logical function on the reconfigurable device is different, multiple logical function modules can be combined into a configuration binary bitstream file of the reconfigurable device.

2: After the system starts working, with the dynamic allocation of tasks, when the logical function block required by the task to be run by a reconfigurable device of a node is not available on the reconfigurable device, the reconfigurable device will send to its own node The host, that is, the general-purpose computer sends a reconfiguration request;

3: After the host receives the reconfiguration request sent by the reconfigurable device of its own node, the host calls the program to reconfigure the reconfigurable device of its own node. The host computer downloads the binary configuration file to the reconfigurable device through the PCI interface.

Step 5: Execution of the task:

Execute the task after the task is assigned, and return the execution result after the task execution ends.

1: If the task is executed on a general-purpose computer, after the execution of the task is completed, the execution result will be returned to the monitoring process, and then the monitoring process will return the result to the calling process of the task.

2: If the task is executed on the reconfigurable device, after the task is executed, the state of the task, including the execution result, is sent to the control module SuperBlock of the reconfigurable device.

3: When SuperBlock receives the task execution result, check the record table of the task. If the task is called by the local general-purpose computer, then send the execution result to the monitoring process of the general-purpose computer through the PCI interface, and then the monitoring process will The result is returned to the task's calling process. If the task is invoked by a remote node, the execution result will be sent through the network to the control module SuperBlock of the reconfigurable device that forwards the task to the neighbor node of the local node.

Claims (4)

1. method that based on network adaptive reconfigurable processing unit array system quickens is characterized in that:

1) task division of original program:

System is divided into a program set of the task of energy independent operating;

2) build based on network adaptive reconfigurable processing unit array:

Based on network adaptive reconfigurable processing unit array is made up of n node, n ∈ [1,2,3...), each node has 1-4 neighbor node that directly is connected by Ethernet, each node is made up of two parts, and first is an all-purpose computer, and second portion is restructural equipment NetFPGA;

According to will on restructural equipment, carrying out of task, initialization restructural equipment;

3) distribution of task:

For the node in any one based on network adaptive reconfigurable processing unit array, oneself be local node, all the other nodes are remote node;

The distribution of task is distributed to the task in the set of tasks in the program computational resource of local node and the computational resource of remote node exactly;

4) change of restructural functions of the equipments:

For the restructural equipment of any one node in n the node, if the restructural device resource free time, and the logic function module wanted of the required by task that is assigned to is currently oneself not possess, and restructural equipment reconfigures functional module so;

5) task executions:

After Task Distribution, execute the task, and after task is carried out end, return execution result.

2. the method that a kind of based on network adaptive reconfigurable processing unit array system according to claim 1 quickens is characterized in that, describedly builds based on network adaptive reconfigurable to put the pe array step as follows:

1) the adaptive reconfigurable processing unit array of being built is based on the high-speed local area network network of 100Mb or 1000Mb transmission rate;

2) restructural equipment NetFPGA has 4 Ethernet interfaces, so the restructural equipment of a node can directly be connected by the high-speed local area network network with maximum four other nodes, communicate, direct-connected node is called neighbor node, and node can communicate by neighbor node and non-neighbor node;

3) connected mode of two of a node part all-purpose computers and restructural equipment is to be connected by the peripheral component interconnection pci interface;

4) the control module SuperBlock on the customization restructural equipment;

5) functional module on the initialization restructural equipment.

3. the method that a kind of based on network adaptive reconfigurable processing unit array system according to claim 1 quickens is characterized in that the change step of described restructural functions of the equipments is as follows:

1) before whole based on network adaptive reconfigurable processing unit array system is started working,, generates this part of configuration to restructural equipment according to will on restructural equipment, moving of task;

2) after system starts working, dynamic assignment along with task, the desired logic function block of task that the restructural equipment of any one node will move in n node does not possess on restructural equipment, and then restructural equipment sends the request of reshuffling to the main frame of own node;

3) after main frame was received the reconfiguration request that the restructural equipment of own node sends, the host computer invokes program was reshuffled the restructural equipment of own node.

4. the method that a kind of based on network adaptive reconfigurable processing unit array system according to claim 2 quickens, it is characterized in that the control module SuperBlock on the described customization restructural equipment serves as the function of the communication controler of any one intra-node restructural equipment and all-purpose computer in n the node; Serve as the function of the restructural equipment and the communication controler between the neighbor node restructural equipment of any one node in n the node; For in system's running, the Task Distribution that restructural equipment is received manages, task is handled, and the residing state of logger task.

Images