Checkpointing for a hybrid computing node

1. A method for checkpointing in a hybrid computing node, the method comprising:

executing a task in a processing accelerator of the hybrid computing node;

creating a checkpoint in a local memory of the processing accelerator, the checkpoint comprising state data to restart execution of the task in the processing accelerator upon a restart operation;

resuming execution of the task in the processing accelerator after creating the checkpoint; and

transferring the state data of the checkpoint from the processing accelerator to a main processor of the hybrid computing node while the processing accelerator is executing the task, wherein the transferring is performed asynchronously on a lower bandwidth interface with the main processor, and a higher speed interface is used to create the checkpoint.

2. The method of claim 1 , wherein execution of the task interacts with accelerator memory space in the local memory, and the checkpoint is stored to a separate memory range in the local memory.

3. The method of claim 1 , wherein the checkpoint is created periodically based on a time interval.

4. The method of claim 1 , wherein the hybrid computing node further comprises multiple processing accelerators that create the checkpoint synchronously in parallel and perform the transferring of the state data to the main processor asynchronously.

5. The method of claim 1 , further comprising:

transferring the state data of the checkpoint from the main processor to another memory system or a network for storage.

6. The method of claim 1 , wherein the processing accelerator is one of a general processing unit or an active memory device.

7. A hybrid computing node comprising:

a main processor; and

a plurality of processing accelerators that create checkpoints synchronously in parallel and perform the transferring of state data to the main processor asynchronously, each of the processing accelerators comprising a local memory and one or more compute engines configured to perform a method comprising:

executing a task in the processing accelerator of the hybrid computing node;

creating a checkpoint in the local memory of the processing accelerator, the checkpoint comprising the state data to restart execution of the task in the processing accelerator upon a restart operation;

resuming execution of the task in the processing accelerator after creating the checkpoint; and

transferring the state data of the checkpoint from the processing accelerator to the main processor of the hybrid computing node while the processing accelerator is executing the task.

8. The hybrid computing node of claim 7 , wherein execution of the task interacts with accelerator memory space in the local memory, and the checkpoint is stored to a separate memory range in the local memory.

9. The hybrid computing node of claim 7 , wherein the checkpoint is created periodically based on a time interval.

10. The hybrid computing node of claim 9 , wherein the transferring is performed asynchronously on a lower bandwidth interface with the main processor, and a higher speed interface is used to create the checkpoint.

11. The hybrid computing node of claim 7 , wherein the processing accelerator is further configured to perform the method comprising:

transferring the state data of the checkpoint from the main processor to another memory system or a network for storage.

12. The hybrid computing node of claim 7 , wherein the processing accelerator is one of a general processing unit or an active memory device.

13. A computer program product for checkpointing in a hybrid computing node, the computer program product comprising:

a tangible storage medium that is not a signal and is readable by a processing circuit and storing instructions for execution by the processing circuit to perform a method comprising:

executing a task in a processing accelerator of the hybrid computing node;

creating a checkpoint in a local memory of the processing accelerator, the checkpoint comprising state data to restart execution of the task in the processing accelerator upon a restart operation;

resuming execution of the task in the processing accelerator after creating the checkpoint; and

transferring the state data of the checkpoint from the processing accelerator to a main processor of the hybrid computing node while the processing accelerator is executing the task, wherein the transferring is performed asynchronously on a lower bandwidth interface with the main processor, and a higher speed interface is used to create the checkpoint.

14. The computer program product of claim 13 , wherein execution of the task interacts with accelerator memory space in the local memory, and the checkpoint is stored to a separate memory range in the local memory.

15. The computer program product of claim 13 , wherein the checkpoint is created periodically based on a time interval.

16. The computer program product of claim 13 , wherein the hybrid computing node further comprises multiple processing accelerators that create the checkpoint synchronously in parallel and perform the transferring of the state data to the main processor asynchronously.

17. The computer program product of claim 13 , wherein the instructions for execution by the processing circuit to perform the method further comprise transferring the state data of the checkpoint from the main processor to another memory system or a network for storage.