Shared memory multiprocessing system employing mixed broadcast snooping and directory based coherency protocols

A multiprocessing system comprising:

• a plurality of clients;
• wherein at least two of the plurality of clients comprise processing subsystems, each including a cache memory;
• and
• wherein at least one of the plurality of clients comprises a memory subsystem including a directory;
• and
• a network separate from said clients;
• wherein said network includes a switch comprising a plurality of ports, wherein each of said plurality of clients may be coupled to a respective one of said plurality of ports, wherein the switch is configured to:
• receive a first coherency request from a given client, wherein said first coherency request is not visible to any other of said plurality of clients;
• and
• transmit a second coherency request in response to receiving said first coherency request;
• wherein the switch further comprises a coherency mode storage unit configured to store an indication to control whether the switch will transmit said second coherency request according to a directory protocol or a broadcast protocol;
• wherein said first coherency request corresponds to a first portion of a region of memory owned by a first client;
• and a third coherency request corresponds to a second portion of the region of memory owned by the first client;
• wherein a first indication stored in the coherency mode storage unit corresponding to the first portion of the region of memory owned by the first client indicates that the switch will transmit the second coherency request according to a directory protocol;
• and
• wherein a second indication stored in the coherency mode storage unit corresponding to the second portion of the region of memory owned by the first client indicates that the switch will transmit a fourth coherency request according to a broadcast protocol.

A multiprocessing system comprising:

• a plurality of clients;
• wherein at least one of the plurality of clients comprises a memory subsystem including a directory, the directory including a plurality of directory entries corresponding to memory locations mapped to said memory subsystem;
• and
• wherein at least two of the plurality of clients comprise processing subsystems, each including a cache for storing blocks corresponding to selected ones of said memory locations;
• and
• a network-separate from said clients, wherein the network comprises:
• a switch comprising a plurality of ports, wherein each of said plurality of clients may be coupled to a respective one of said plurality of ports;
• and
• a coherency mode storage unit configured to store an indication of whether a given block is a point-to-point mode block or a broadcast mode block;
• wherein the switch is configured to route a coherency request for said given block initiated by one of said processing subsystems to said directory in response to said indication indicating said given block is a point-to-point mode block, and wherein said network is configured to broadcast said coherency request to said memory subsystem and to each of said plurality of processing subsystems in response to said indication indicating said given block is a broadcast mode block.

A switch comprising:

• a plurality of ports, wherein each of said ports is configured to be coupled to one of a plurality of clients which are separate from said switch, wherein at least two of the plurality of clients comprise processing subsystems, each including a cache memory;
• and wherein at least one of the plurality of clients comprises a memory subsystem including a directory;
• wherein the switch is configured to:
• receive a first coherency request from a given client, wherein said first coherency request is not visible to any other of said plurality of clients;
• and
• transmit a second coherency request in response to receiving said first coherency request;
• wherein said first coherency request corresponds to a first portion of a region of memory owned by a first client;
• and a third coherency request corresponds to a second portion of the region of memory owned by the first client;
• wherein a first indication stored in the coherency mode storage unit corresponding to the first portion of the region of memory owned by the first client indicates that the switch will transmit the second coherency request according to a directory protocol;
• and
• wherein a second indication stored in the coherency mode storage unit corresponding to the second portion of the region of memory owned by the first client indicates that the switch will transmit a fourth coherency request according to a broadcast protocol.

A multiprocessing system comprising:

• a plurality of clients;
• wherein at least two of the plurality of clients comprise processing subsystems, each including a cache memory;
• and
• wherein at least one of the plurality of clients comprises a memory subsystem including a directory;
• and
• a network separate from said clients;
• wherein said network includes a switch comprising a plurality of ports, wherein each of said plurality of clients may be coupled to a respective one of said plurality of ports, wherein the switch is configured to:
• receive a first coherency request from a given client, wherein said first coherency request is not visible to any other of said plurality of clients;
• and
• transmit a second coherency request in response to receiving said first coherency request;
• wherein the switch further comprises a coherency mode storage unit configured to store an indication to control whether the switch will transmit said second coherency request according to a directory protocol or a broadcast protocol;
• wherein ownership transitions for a first set of cache lines are governed according to a sequence of actions within said network;
• and
• wherein ownership transitions for a second set of cache lines are governed according to a sequence of actions within said directory;
• wherein if said second coherency request is transmitted by said switch according to said directory protocol, said second coherency request is transmitted by said switch to said memory subsystem, said directory is accessed by said memory subsystem, and responsive coherency commands are provided to one or more of said plurality of processing subsystems dependent upon information contained within said directory.

A multiprocessing system comprising:

• a plurality of clients;
• wherein at least two of the plurality of clients comprise processing subsystems, each including a cache memory;
• and
• wherein at least one of the plurality of clients comprises a memory subsystem including a directory;
• and
• a network separate from said clients;
• wherein said network includes a switch comprising a plurality of ports, wherein each of said plurality of clients may be coupled to a respective one of said plurality of ports, wherein the switch is configured to:
• receive a first coherency request from a given client, wherein said first coherency request is not visible to any other of said plurality of clients;
• and
• transmit a second coherency request in response to receiving said first coherency request;
• wherein the switch includes a control unit configured to monitor bandwidth utilization of said network and to control whether the second coherency request is transmitted through said network according to a directory protocol or a broadcast protocol depending upon the bandwidth utilization.

A method of operating a multiprocessing system including a plurality of clients, wherein at least two of the plurality of clients comprise processing subsystems and wherein at least one of the plurality of clients comprises a memory subsystem;

• and a network separate from said clients, wherein the network comprises a switch comprising a plurality of ports wherein each of said plurality of clients may be coupled to a respective one of said plurality of ports, the method comprising:
• monitoring bandwidth utilization of said network;
• storing an indication in said switch to control whether, in response to receiving a first coherency request, a second coherency request is transmitted through said network according to a directory protocol or a broadcast protocol;
• receiving said first coherency request in said switch from a first processing subsystem, wherein said first coherency request is not visible to any other of said plurality of clients;
• and
• said switch conveying said second coherency request according to said indication;
• wherein said indication is set in dependence upon said bandwidth utilization.