US 6,275,900 B1Granted

Hybrid NUMA/S-COMA system and method

AssigneeInternational Business Machines Corporation

InventorDean A. Liberty

Life of this patent

Estimating the term…

Loading patent…

Abstract

A hybrid non-uniform-memory-architecture/simple-cache-only-memory-architecture (NUMA/S-COMA) memory system and method are described useful in association with a computer system having a plurality of nodes coupled to each other. The plurality of nodes include NUMA memory which are configured to store data lines. The NUMA memories include a NUMA coherence subsystem for coordinating transfer of data between the nodes. At least one S-COMA cache is provided on at least one node of the computer system. The at least one S-COMA cache is configured to employ the NUMA coherence subsystem in sending data communication to or receiving data communication from another node of the plurality of nodes of the computer system. Data stored at another node of the system is accessed using a home node real address as the network address. The home node real address is translated into a local real address at the client node using a boundary function translation table. The NUMA coherence subsystem is employed by the S-COMA cache to provide data reference capture, data movement and coherence mechanisms, thereby avoiding the need for a separate S-COMA coherence mechanism to accomplish these functions.

Claims18

Dependent claims

A hybrid non-uniform memory architecture/simple cache-only-memory-architecture (NUMA/S-COMA) memory system useful in association with a computer system having a plurality of nodes coupled to each other, and wherein data is stored in said hybrid NUMA/S-COMA memory system as a plurality of pages, each said page including at least one data line, said hybrid NUMA/S-COMA memory system comprising:

a plurality of NUMA memories configured to store said at least one data line, each NUMA memory of said plurality of NUMA memories residing at a different node of said plurality of nodes of said computer system, and wherein said plurality of NUMA memories includes a NUMA coherence subsystem for coordinating transfer of data therebetween;
at least one S-COMA cache configured to store at least one of said plurality of pages, each S-COMA cache of said at least one S-COMA cache residing at a different node of said plurality of nodes or said computer system, and wherein said at least one S-COMA cache employs said NUMA coherence subsystem in sending data communication to or receiving data communication from another node of said plurality of nodes of the computer system;
wherein said at least one S-COMA cache comprises a plurality of S-COMA caches, and wherein said NUMA coherence subsystem includes a NUMA coherence protocol on a client node of said plurality of nodes caching data in one of said S-COMA caches from a home node of said plurality of nodes of said computer system, said NUMA coherence protocol being employed to send S-COMA coherence messages between said client node and said home node;
and
wherein said NUMA coherence subsystem with said coherence protocol at said client node comprises a boundary function translation table at said client node for transitioning between a home node real address and a client node real address when sending a message between said one S-COMA cache of said client node and said home node, wherein said home node real address comprises a network address employed by said NUMA coherence subsystem in transferring data between said home node and said S-COMA cache of said client node, and wherein said home node real address comprises a local physical address of the home node without need of translating between global addresses and real addresses on the home node.

2 more elements

A method for communicating data between a client node and a home node of a computer system having a plurality of nodes coupled to each other, wherein said computer system employs a hybrid non-uniform-memory architecture/simple-cache-only-memory-architecture (NUMA/S-COMA) memory system having a plurality of NUMA memories configured to store data, each NUMA memory residing at a different node of said computer system, and at least one S-COMA cache configured to store data, each S-COMA cache residing at a different node of said computer system, said client node including an S-COMA cache of said at least one S-COMA cache, said communication method comprising:

(i) generating at said client node of said plurality of nodes of said computer system a data request with a real address of memory;
(ii) determining whether said real address comprises a local real address at said client node;
(iii) when said real address comprises a local real address, determining whether a boundary function translation is required of said local real address to translate said local real address into a home node real address;
and
(iv) when said boundary function translation is required, translating said local real address into said home node real address, wherein said home node real address comprises a network address to be employed by said client node in requesting access to data at said home node real address, and wherein said home node real address comprises a local physical address of the home node of said plurality of nodes without need of translating between global addresses and real addresses on the home node.

2 more elements

An article of manufacture, comprising:

at least one computer usable medium having computer readable program code means embodied therein for causing communicating of data between a client node and a home node of a computer system having a plurality of nodes coupled to each other, wherein said computer system employs a hybrid non-uniform-memory architecture/simple-cache-only-memory-architecture (NUMA/S-COMA) memory system having a plurality of NUMA memories configured to store data, each NUMA memory residing at a different node of said computer system, and at least one S-COMA cache configured to store data, each S-COMA cache residing at a different node of said computer system, said client node including an S-COMA cache of said at least one S-COMA cache, the computer readable program code means in the article of manufacture comprising:
(i) computer readable program code means for causing a computer to effect generating at said client node of said plurality of nodes a data request with a real address of memory;
(ii) computer readable program code means for causing a computer to effect determining whether said real address comprises a local real address at said client node;
(iii) computer readable program code means for causing a computer to effect determining whether a boundary function translation is required when said real address comprises a local real address, said boundary function translation translating said local real address into a home node real address;
and
(iv) computer readable program code means for causing a computer to effect translating said local real address into said home node real address when said boundary function translation is required, wherein said home node real address comprises a network address to be employed by said client node in requesting access to data at said home node real address, and wherein said home node real address comprises a local physical address of the home node of said plurality of nodes without need of translating between global addresses and real addresses on the home node.

3 more elements

Dependencies are inferred from the claim text.