Adaptively Handling Remote Atomic Execution

• US 20100332801A1
• Filed: 06/26/2009
• Published: 12/30/2010
• Est. Priority Date: 06/26/2009
• Status: Active Grant

• Current Assignee: Intel Corporation
• Original Assignee: Intel Corporation
• Inventors: Sprangle, Eric, Mejia, John, Sundararaman, Ramacharan, Rajwar, Ravi, Forsyth, Andrew Thomas, Fryman, Joshua B., Espasa, Roger, Juan, Toni, Grochowski, Edward T.
• Granted Patent: US 8,533,436 B2
• Time in Patent Office: Days
• US Class Current: 712/208
• CPC Class Codes: G06F 12/0817 using directory methods
• CPC Class Codes: G06F 9/3004 to perform operations on memory
• CPC Class Codes: G06F 9/30087 Synchronisation or serialisation instructions
• CPC Class Codes: G06F 9/30185 according to one or more bits in the instruction, e.g. prefix, sub-opcode
• CPC Class Codes: G06F 9/3834 Maintaining memory consistency
• CPC Class Codes: G06F 9/3861 Recovery, e.g. branch miss-prediction, exception handling error detection or correction G06F11/00

Abstract

In one embodiment, a method includes receiving an instruction for decoding in a processor core and dynamically handling the instruction with one of multiple behaviors based on whether contention is predicted. If no contention is predicted, the instruction is executed in the core, and if contention is predicted data associated with the instruction is marshaled and sent to a selected remote agent for execution. Other embodiments are described and claimed.

Claims

1. A method comprising:

receiving a first instruction for decoding in a first core of a processor and determining whether contention is predicted with respect to the first instruction; and

dynamically handling the first instruction with one of multiple behaviors, including if no contention is predicted executing the first instruction in the first core, and if contention is predicted marshaling data associated with the first instruction in the first core and sending a remote execution request with the marshaled data to a selected remote agent for execution of the first instruction.

2. The method of claim 1, wherein execution of the first instruction in the first core or the selected remote agent is transparent to a user and an operating system.

3. The method of claim 1, further comprising determining whether to execute the first instruction in the first core or the selected remote agent at decode of the first instruction based on the prediction.

4. The method of claim 1, further comprising ignoring the contention prediction if the first instruction is not of a selected type.

5. The method of claim 4, further comprising sending a request from the first core to a directory to determine an owner of data associated with the first instruction, the request to indicate to the directory that the first core is to request the remote execution, wherein the directory is to update an entry of the directory associated with the data to indicate access of the data by the first core and to update an in-flight indicator to denote that the remote execution request is pending.

6. The method of claim 5, further comprising receiving a message from the directory in the first core to indicate an owner core for the data.

7. The method of claim 6, further comprising sending the remote execution request to the owner core for the data from the first core, responsive to the message from the directory.

8. The method of claim 5, wherein updating the directory entry comprises setting a first indicator of a core field associated with the first core of a vector to indicate access of the data via an atomic operation.

9. The method of claim 8, wherein when a timer expires, determining in the directory whether a plurality of indicators of core fields of the vector are set and if so, updating a prediction feedback value to reflect contention.

10. The method of claim 9, further comprising receiving in the first core a result of the remote execution request and an update for the prediction, and updating an entry of a predictor of the first core corresponding to the first instruction with the update.

11. An apparatus comprising:

a multi-core processor including a plurality of cores, each of the cores including a front end to decode an instruction and determine whether contention is predicted with respect to the instruction, an execution logic to execute the instruction locally if no contention is predicted, and if contention is predicted, to marshal data associated with the instruction and send a remote execution request with the marshaled data to a selected remote agent for execution of the instruction, and a predictor having a plurality of entries each to store a prediction as to whether an instruction associated with the entry is to be contended during execution.

12. The apparatus of claim 11, further comprising a directory having a plurality of entries each to store tag information regarding a cacheline present in one or more cores of the multi-core processor, each entry further including a contention vector having a plurality of core fields each associated with a core, each core field including a first indicator and a second indicator each to indicate access of the cacheline via an atomic operation by the associated core within a time period.

13. The apparatus of claim 12, wherein each directory entry further includes a tag portion to store an in-flight indicator to indicate that a remote execution request associated with the cache line is pending.

14. The apparatus of claim 12, wherein the core is to send a first message to the directory to determine an owner of data associated with the instruction and to indicate that the core is to request a remote execution.

15. The apparatus of claim 14, wherein the core is, responsive to receipt of a response message from the directory, to send a second message to an owner of the data, the second message including the remote execution request and the marshaled data.

16. The apparatus of claim 15, wherein the core is to receive a result of the remote execution request and an update for the prediction, and update an entry of the predictor corresponding to the instruction with the update.

17. The apparatus of claim 11, wherein the selected remote agent is a memory controller of the multi-core processor, the memory controller including an execution logic.

18. A system comprising:

a processor including having a front end to decode an instruction and determine whether contention is predicted with respect to the instruction, a dynamic execution logic to execute the instruction locally if no contention is predicted and otherwise to marshal data associated with the instruction and send a remote execution request with the marshaled data to a selected remote agent for execution of the instruction, and a predictor having a plurality of entries each to store a prediction as to whether an instruction associated with the entry is to be contended during execution;

a directory having a plurality of entries each to store tag information regarding a cacheline stored in a cache of the system, each entry further including a contention vector having a plurality of core fields each associated with a core, each core field including a first indicator and a second indicator each to indicate access of the cacheline via an atomic operation by the associated core within a time period; and

a dynamic random access memory (DRAM) coupled to the directory and the processor.

19. The system of claim 18, wherein the processor is to send a first message to the directory to determine an owner of data associated with the instruction and to indicate that the processor is to request a remote execution.

20. The system of claim 19, wherein the processor, responsive to receipt of a response message from the directory, is to send a second message to the selected remote agent corresponding to an owner of the data, the second message including the remote execution request and the marshaled data, and wherein the processor is to receive a result of the remote execution request and an update for the prediction, and update an entry of the predictor corresponding to the instruction with the update.