Noc arbitration method based on global node information in gpgpu

Abstract

The invention discloses a NoC arbitration method based on global node information in a GPGPU. The method comprises the following steps: S1. collecting performance information of each computing node inthe network, setting the priority of each computing node according to the performance information, and broadcasting the priority to all computing nodes for global synchronous updating; S2. when the computing node issues a memory access request, grouping data packets that need to be injected into the network in each computing node to obtain grouping information of the data packets; when the computing node performs requesting arbitration, performing inter-group arbitration according to the grouping information, and performing intra-group arbitration according to the priority of the computing node, and multiple requests win after the intra-group arbitration, performing additional arbitration by using a local arbitration mechanism. The NoC arbitration method disclosed by the invention has theadvantages of simple implementation method, high arbitration efficiency, fairness of arbitration, optimization of resource allocation, consistency of arbitration results, and improvement of system performance.

Description

technical field [0001] The present invention relates to the field of network-on-a-chip (NoC) arbitration technology in GPGPU (General Purpose Graphic Process Unit, general purpose graphics processor), in particular to a NoC arbitration method based on global node information in GPGPU. Background technique [0002] GPGPU uses a large number of simple processor cores to obtain high throughput by processing large amounts of data in parallel, and switches between a large number of tasks to hide the memory access delay caused by off-chip memory access. Due to the simple structure of the processor core, a large number of transistors can be integrated on the GPGPU, which makes it unique in single-precision and double-precision floating-point operations. The super computing power in floating-point computing makes GPGPU widely used in the field of high-performance computing and becomes an important means to solve computing-intensive scientific and engineering problems. While the hig...

AI Technical Summary

Problems solved by technology

[0006] 3) The read request message in the request network is small, and the write request message is large but small in number; and there are a large number of read request reply messages in the reply network, and these messages usually carry the read data, so the message Large quantity and long length. The above communication characteristics make the request network load low and the reply network load high. The unbalanced load is also one of the reasons why the existing on-chip network structure cannot give full play to its performance advantages.

[0007] 4) The communication between different computing nodes in a general-purpose processor is mainly caused by the shared storage programming model. In the shared storage mode, in order to improve memory access efficiency and system performance, different computing nodes have private caches, and data at the same address is in There may be different copies in different caches, so there is a cache consistency problem

[0010] ②Round-Robin arbitrator adopts polling arbitration mechanism: this type of arbitrator is obtained from the input port or virtual channel that sends the memory access request in the current cycle, and its priority is the lowest in the next round of arbitration after being satisfied, which is different from the above fixed priority Compared with arbitration, this type of round-robin scheduling method increases fairness, but the performance of the entire system is not high;

[0011] ③The Matrix arbitrator stores the elements of the priority relationship between N requests in the form of an N x N matrix, and the element W in the matrix ij Indicates the priority relationship arbitration between the input port i and the input port j that sends the request. Compared with polling arbitration, this type of arbitration has higher data throughput and faster operation processing speed, but its high-speed operation ability is At the cost of taking up a lot of resources and area overhead, and the performance of the entire system is still not high

[0012] In addition to the above-mentioned types of arbitrators, there are also arbitrators for on-chip multi-core networks or long-path network designs. However, the current NoC arbitration mechanisms mentioned above are all local arbitration and random scheduling, and the design goals are all focused on ensuring fairness in arbitration. In terms of improving service quality, the arbitration mechanism does not fully consider and utilize global information such as computing nodes and network traffic, and ignores the role of the arbitration mechanism in improving overall system performance. Although the fairness of arbitration can be guaranteed to a certain extent, due to local Random arbitration is only scheduled based on the memory access request information received at the current node. The design of the arbitration mechanism does not fully consider the characteristics of all memory access requests sent by each computing node, and the matching degree between the arbitration result and the program requirements is low; at the same time, each router independently Scheduling may lead to incoordination, mutual interference and even contradictions among the scheduling results. Therefore, the current NoC arbitration mechanism cannot improve the overall system performance, and may even degrade the system performance.

Images