Method for building multi-processor system with nodes having multiple cache coherency domains
A method for building a multi-processor system with nodes having multiple cache coherency domains. In this system, a directory built in a node controller needs to include processor domain attribute information, and the information can be acquired by configuring cache coherency domain attributes of ports of the node controller connected to processors. In the disclosure herein, the node controller can support the multiple physical cache coherency domains in a node.
1. A method for building a multi-processor system with nodes having multiple cache coherency domains, wherein in the multi-processor node system, a unified logical cache coherency domain is built in a controller of each node, and several processors and the node controller form multiple physical cache coherency domains isolated from one another; comprising a cache coherency domain building manner, a cache coherency domain dividing manner, and a directory structure manner, wherein
(1) the building manner of the multiple physical cache coherency domains comprises:
1) connecting first-level nodes directly through node controllers or through a node router, so as to form a single cache coherency domain of the first-level nodes;
2) directly connecting a second-level node controller to the first-level node controllers, and building a unified logical cache coherency domain in the second-level node controller, wherein the cache coherency domain comprises multiple physical cache coherency domains that are isolated from one another and formed by the several first-level node controllers and the second-level node controller;
3) connecting second-level nodes directly through second-level node controllers or through a node router, an as to form a single cache coherency domain of the second-level nodes; and
4) extending to n th (n>2)-level nodes based on the manners of 1)-3), so as to implement a multi-processor system with more levels of nodes;
(2) the cache coherency domain dividing manner comprises:
1) for the first-level node controller, using information of various ports connected to the processors as a domain attribute information source of the physical cache coherency domains formed by the several processors and the node controller, identifying, by each port domain information register, a physical sub-domain of the port, and configuring domain attribute information of each port by the port domain information register according to system requirements, thereby implementing division of various physical cache coherency domains formed by the several processors and the node controller;
2) for the second-level node controller, using information of various ports connected to the first-level node controllers as a domain attribute information source of the physical cache coherency domains formed by the several first-level node controllers and the second-level node controller, and configuring domain attribute information of each port according to the system requirements, thereby implementing division of various physical cache coherency domains formed by the several first-level node controllers and the second-level node controller; and
3) configuring port domain attributes of an n th (n>2)-level node controller by using the configuration manner of port domain attributes of the second-level node controller described in item 2) of the dividing manner;
(3) the directory structure manner comprises:
1) the first-level node controller using a local and remote double directory structure, wherein for a remote data directory, directory entries record processors possessing a unified data copy in the local node, as well as domain attribute information and coherence state information of the processors; and for a local data directory, directory entries record remote nodes possessing the data copy and coherence state information of the remote nodes, and a flag bit is set to indicate whether other sub-domains in the local node other than a physical sub-domain where the data is located possess the data copy;
2) the second-level node controller using a local and remote double directory structure, wherein for a remote data directory, directory entries record first-level nodes possessing a unified data copy in the local second-level node, as well as domain attribute information and coherence state information of the first-level nodes; and for a local data directory, directory entries record remote second-level nodes possessing the unified data copy and coherence state information of the remote second-level nodes, and a flag bit is set to indicate whether other sub-domains in the local second-level node other than a physical sub-domain where the data is located possess the data copy;
3) configuring a directory structure of an n th (n>2)-level node controller by using the directory structure manner of the second-level node controller described in item 2) of the directory structure manner;
4) the first-level node controller using a local/remote unified directory structure, where directory entries record all processors in the system other than the processors of this physical sub-domain possess unified data copy information and coherence state information;
5) the second-level controller using a local/remote unified directory structure, where directory entries record all first-level node controllers in the system other than the first-level node controllers of this physical sub-domain possess unified data copy information and coherence state information; and
6) configuring a directory structure of an n th (n>2)-level node controller by using the directory structure manner of the second-level node controller described in item 5) of the directory structure manner.