Method for supporting improved burst transfers on a coherent bus
View Patent ↗In a multiprocessor system, comprising master and slave processors, a cache coherency controller, and address concentration devices; a method for improving coherent data transfers is described. A command transaction is generated, and a subsequent command from an initiator. Tags added to the responses or further request responses, stream on high-speed busses. Snoops and accumulated snoops expand on cacheline requests as each processor separates burst commands into multiple cacheline requests. Address concentrators containing a cacheline queue function, funnel transaction requests to a global serialization device, where a queuing process prioritizes indicia and coordinates the results among the processors. The cache issues a single burst command for each affected line. System coherency, performance, and latency improvements occur. Additional support for burst transfers between coherent processors is provided.
1. A method in a multiprocessor system, including at least two processors and a cache coherency controller, coupled to address concentration devices, the method operable in said cache coherency controller for improving coherent data transfers, the method comprising the steps of:
initiating a first memory transaction request command to a first processor of said multiple processors;
determining priority receipt status of a next memory transaction request;
expanding snoop responses and accumulated snoop responses to provide a coherency action for all cacheline requests utilizing a burst command;
forwarding said transaction requests from said master or said controller to a solitary global serialization device, said serialization device further comprising a multiple cacheline request indicator;
grouping multiple and sequential coherent transfers into a burst operation on said processor bus;
determining occurrence of the burst operation;
timing said burst operation;
snooping said burst operation to form a plurality of snoop replies;
concentrating all addresses of said plurality of snoop replies to form a combined snoop response; and
broadcasting the combined snoop response to a plurality of device entities.
2. The method of claim 1 wherein said determining said occurrence of said burst operation comprises detecting an indication that a plurality of cachelines is requested.
3. The method of claim 2 wherein each processor of said at least two processors supports snooping a plurality of cachelines for the burst operation.
4. The method of claim 1 wherein said snooping said burst operation comprises separating said burst operation into multiple cacheline requests on a bus.
5. The method of claim 1 wherein said snooping said burst operation comprises at least a bus and at least a cache directly supporting a burst operation.
6. The method of claim 1 wherein broadcast of said combined snoop response is between pluralities of processors.
7. The method of claim 1 wherein said determining said occurrence of said burst operation includes the step of:
determining whether a previous cacheline request corresponding to said processor is already present in a first-in first-out queue; and
queuing said cacheline request in response to a determination that said cacheline request is not already present in said first-in first-out queue.
8. The method of claim 1 wherein accumulated snoop responses agglomerate as a combined response only on notification of individual responses of said entities within said processors.
9. A method of performing a burst transfer on a bus with cache coherency, the method comprising:
initiating a burst transfer command on a bus connecting a plurality of processors, wherein the burst transfer command is a request to access a block of data that comprises two or more cachelines;
performing, by each given processor within the plurality of processors, a cache snoop of the burst transfer command to form a burst snoop response;
accumulating burst snoop responses for the two or more cachelines from the plurality of processors to form a combined snoop response; and
broadcasting the combined snoop response to the plurality of processors,
wherein performing a cache snoop of the burst transfer command comprises:
splitting, by a snooping processor within the plurality of processors, the burst transfer command into a plurality of single cacheline requests;
forwarding the plurality of single cacheline requests to a plurality of entities within the snooping processor; and
performing, by each entity within the snooping processor, a cache snoop of each of the plurality of single cacheline requests to form a plurality of single cacheline snoop responses.
10. The method of claim 9 , wherein initiating a burst transfer command comprises:
responsive to a plurality of transfer commands requesting to access a plurality of sequential cachelines from a plurality of entities within an initiating processor, accumulating the plurality of transfer commands to form the burst transfer command.
11. The method of claim 10 , further comprising:
responsive to receipt of the combined snoop response at the initiating processor, separating the combined snoop response into a plurality of individual snoop responses; and
forwarding the plurality of individual snoop responses to the pluralities of entities within the initiating processor.
12. The method of claim 9 , wherein performing a cache snoop of the burst transfer command further comprises:
returning the plurality of single cacheline snoop responses from each entity within the snooping processor to a gathering function within the snooping processor, wherein the gathering function accumulates the plurality of single cacheline snoop responses to form a burst snoop response for the snooping processor.
13. A data processing system comprising:
a plurality of processors; and
a bus connecting the plurality of processors,
wherein an initiating processor from the plurality of processors initiates a burst transfer command on the bus and wherein the burst transfer command is a request to access a block of data that comprises two or more cachelines;
wherein each given processor within the plurality of processors performs a cache snoop of the burst transfer command to form a burst snoop response;
wherein a snoop response gathering function accumulates burst snoop responses for the two or more cachelines from the plurality of processors to form a combined snoop response and broadcasts the combined snoop response to the plurality of processors,
wherein a separation function within a snooping processor within the plurality of processors splits the burst transfer command into a plurality of single cacheline requests and forwards the plurality of single cacheline requests to a plurality of entities within the snooping processor; and
wherein each entity within the snooping processor performs a cache snoop of each of the plurality of single cacheline requests to form a plurality of single cacheline snoop responses.
14. The data processing system of claim 13 , wherein, responsive to a plurality of transfer commands requesting to access a plurality of sequential cachelines from a plurality of entities within the initiating processor, a gathering function within the initiating processor accumulates the plurality of transfer commands to form the burst transfer command.
15. The data processing system of claim 14 , wherein, responsive to receipt of the combined snoop response at the initiating processor, a separation function within the initiating processor separates the combined snoop response into a plurality of individual snoop responses and forwards the plurality of individual snoop responses to the pluralities of entities within the initiating processor.
16. The data processing system of claim 13 , wherein a gathering function within the snooping processor accumulates the plurality of single cacheline snoop responses to form a burst snoop response for the snooping processor.