High-performance streaming of ordered write stashes to enable optimized data sharing between I/O masters and CPUs

• US 10,452,593 B1
• Filed: 07/05/2018
• Issued: 10/22/2019
• Est. Priority Date: 05/03/2018
• Status: Active Grant

• Current Assignee: Arm Ltd. (United Kingdom) (SoftBank Group Corp.)
• Original Assignee: Arm Ltd. (United Kingdom) (SoftBank Group Corp.)
• Inventors: Jalal, Jamshed, Ringe, Tushar P., Tummala, Ashok Kumar, Ramagiri, Gurunath
• Primary Examiner(s): Park, Ilwoo
• Application Number:
  US16/027,490
• Publication Number:
  US 20190340147A1
• Time in Patent Office: 474 Days
• Field of Search: None
• CPC Class Codes: G06F 12/0813 with a network or matrix configuration
• CPC Class Codes: G06F 12/0831 using a bus scheme, e.g. with bus monitoring or watching means
• CPC Class Codes: G06F 13/382 using universal interface adapter
• CPC Class Codes: G06F 13/4217 with synchronous protocol
• CPC Class Codes: G06F 15/17337 Direct connection machines, e.g. completely connected computers, point to point communication networks coupling between buses G06F13/4004
• CPC Class Codes: G06F 15/1735 Network adapters, e.g. SCI, Myrinet protocol engines H04L69/12
• CPC Class Codes: G06F 2212/1016 Performance improvement

Abstract

A data processing network and method of operation thereof are provided for efficient transfer of ordered data from a Request Node to a target node. The Request Node send write requests to a Home Node and the Home Node responds to a first write request when resources have been allocated the Home Node. The Request Node then sends the data to the written. The Home Node also responds with a completion message when a coherency action has been performed at the Home Node. The Request Node acknowledges receipt of the completion message with a completion acknowledgement message that is not sent until completion messages have been received for all write requests older than the first write request for the ordered data, thereby maintaining data order. Following receipt of the completion acknowledgement for the first write request, the Home Node sends the data to be written to the target node.

Claims

1. A method for transferring ordered data from a source device to a target device in a data processing network, where the source device is coupled to a Request Node of the data processing network, the method comprising:

receiving, by the Request Node, a first write request from the source device, where the first write request is targeted to the target device and is associated with a stream identifier;

sending, by the Request Node to a Home Node of the data processing network, a second write request;

sending, by the Home Node to the Request Node responsive to the second write request, a write data buffer credit message, identifying a data buffer for receiving data to be written, and a completion message indicating completion by the Home Node;

sending, by the Request Node to the Home Node responsive to the write data buffer credit message, the data to be written;

sending, by the Request Node to the Home Node responsive to the completion message, a first completion acknowledgement message;

sending, by the Home Node to the target device responsive to the first completion acknowledgement message, the data to be written; and

sending, by the target device to the Home Node responsive to receiving the data to be written, a second completion acknowledgement message;

where the first completion acknowledgement message is not sent by the Request Node until completion messages for all write requests older than the first write request and having the same stream identifier as the first write request have been received by the Request Node.

2. The method of claim 1, where the second write request indicates to the Home Node that the data to be written is part of a set of ordered data and that observation order is to be controlled by release of a completion acknowledgement message from the Request Node.

3. The method of claim 1, where the write data buffer credit message and the completion message sent from the Home Node to the Request Node are combined as one message when the Home Node determines that no target devices have older copies of the data to be written.

4. The method of claim 1, where the first completion acknowledgement message and the data to be written sent from the Request Node to the Home Node are combined as one message when the Request Node determines that all older writes have received completion messages from the Home Node.

5. The method of claim 1, further comprising:

sending, by the Home Node to the target device, a snoop message in accordance with a coherence protocol; and

sending, by the target device to the Home Node, a response to the snoop message;

where the data to be written is not sent from the Home Node to the target device until the response to the snoop message is received by the Home Node.

6. The method of claim 1, where the first write request comprises a stash request and where the data to be written is stored in a unique-dirty coherence state at the target device.

7. The method of claim 1, further comprising:

detecting, by the Request Node, deadlock resulting from resource conflict in an interconnect of the data processing network;

sending, by the Request Node to the Home Node, a message to cancel the first write request;

sending, by the Home Node responsive to the message to cancel the first write request, a snoop request to a memory controller to retrieve a latest copy of the data associated with an address of the data to be written; and

receiving, by the Home Node from the memory controller, the latest copy of the data associated with the address of the data to be written;

where the latest copy of the data associated with an address of the data to be written is sent to the target device instead of the data to be written.

8. The method of claim 1, where the source device comprises an input/output device, an accelerator device, or a graphics processing unit.

9. The method of claim 1, further comprising the Home Node freeing the data buffer responsive to the second completion acknowledgement message.

10. A method for transferring ordered data of a data stream from a Request Node to a target device in a data processing network, the method comprising:

sending, by the Request Node to one or more Home Nodes of the data processing network, a plurality of first write requests;

for each first write request of the plurality of first write requests;

sending, by a Home Node of the one or more Home Nodes to which the first write request was sent, a write data buffer credit message, identifying a data buffer for receiving data associated with the first write request, and a completion message indicating completion by the Home Node;

sending, by the Request Node to the Home Node responsive to the write data buffer credit message, data to be written;

sending, by the Request Node to the Home Node responsive to the completion message, a first completion acknowledgement message;

sending, by the Home Node to the target device responsive to the first completion acknowledgement message, the data to be written; and

sending, by the target device to the Home Node responsive to receiving the data to be written, a second completion acknowledgement message;

where the first completion acknowledgement message is not sent by the Request Node until completion messages for all older write requests of the data stream have been received by the Request Node.

11. The method of claim 10, where the write data buffer credit message and the completion message sent from the Home Node to the Request Node are combined as one message for each first write request.

12. The method of claim 10, where the first completion acknowledgement message and the data to be written sent from the Request Node to the Home Node are combined as one message for each first write request.

13. The method of claim 10, further comprising:

sending, by the Home Node of the one or more Home Nodes to which the first write request was sent, a snoop message to the target device in accordance with a coherence protocol; and

sending, by the target device to the Home Node, a response to the snoop message;

where data associated with the first write request is not sent from the Home Node to the target device until the response to the snoop message is received by the Home Node.

14. The method of claim 10, where the plurality of first write requests comprise stash requests and where the data associated with the first write requests is stored in a unique-dirty coherence state at the target device.

15. The method of claim 10, further comprising:

detecting, by the Request Node, deadlock resulting from resource conflict in an interconnect of the data processing network;

sending, by the Request Node to a Home Node of the one or more Home Nodes, a message to cancel a first write request of the plurality of first write requests;

sending, by the Home Node responsive to the message to cancel the first write request, a snoop request to a memory controller to retrieve a latest copy of the data associated with an address of the data to be written; and

receiving, by the Home Node from the memory controller, the latest copy of the data associated with the address of the data of the cancelled first write request;

where the latest copy of the data of the cancelled first write request is sent to the target device instead of the data associated with the first write request.

16. The method of claim 10, where the ordered data of the data stream is received at the Request Node from an input/output device, an accelerator device, or a graphics processing unit.

17. A data processing network comprising:

Target Node;

a Request Node configured for receiving a first write request from a source device for data to be written to the Target Node, the data associated with an ordered data stream and the first write request associated with a data stream identifier;

a Home Node that provides a point of coherency for the data processing network, where the Home Node is coupled to the Target Node and the Request Node via an interconnect;

where the Request Node is configured for;

sending a second write request to the Home Node responsive to the first write request from the source device;

receiving, from the Home Node responsive to the second write request, a write data buffer credit message, identifying a data buffer for receiving the data to be written, and a completion message;

sending, to the Home Node responsive to the write data buffer credit message, the data to be written; and

sending, to the Home Node responsive to the completion message, a first completion acknowledgement message, where the first completion acknowledgement message is not sent until completion messages for all write requests that are older than the first write request and are associated with the same data stream identifier as the first write request have been received by the Request Node;

where the Home Node is configured for sending the data to be written to the Target Node responsive to the first completion acknowledgement message; and

where the Target Node is configured for sending a second completion acknowledgement message to the Home Node responsive to receiving the data to be written.

18. The data processing network of claim 17, where the write data buffer credit message and the completion message sent from the Home Node to the Request Node are combined as one message.

19. The data processing network of claim 17, where the first completion acknowledgement message and the data to be written sent from the Request Node to the Home Node are combined as one message.

20. The data processing network of claim 17, where:

the Home Node is configured for sending a snoop message to the Target Node in accordance with a coherence protocol;

the Target Node is configured for sending a response to the snoop message back to the Home Node; and

the data to be written is not sent from the Home Node to the Target Node until the response to the snoop message is received by the Home Node.

21. The data processing network of claim 17, where the second write request comprises a stash request and where the data to be written is stored in a unique-dirty coherence state in a cache of the Target Node.

22. The data processing network of claim 17, where:

the Request Node is further configured for detecting deadlock resulting from resource conflict in an interconnect of the data processing network and sending a message to cancel the first write request to the Home Node; and

the Home Node is further configured, responsive to the message to cancel the first write request, for;

sending a snoop request to a memory controller of the data processing network to retrieve a latest copy of the data associated with an address of the data to be written;

receiving, from the memory controller, the latest copy of the data associated with the address of the data to be written; and

sending the latest copy of the data associated with an address of the data to be written to the Target Node instead of the data to be written.

23. The data processing network of claim 17, where the source device comprises an input/output device, an accelerator device, or a graphics processing unit.

24. A non-transitory computer readable medium comprising instructions of a Hardware Description Language descriptive of the data processing network of claim 17.

25. A non-transitory computer readable medium storing a netlist of components and connectivity of the data processing network of claim 17.