Userspace split data-centric heterogeneous computing architecture
One example method includes transmitting, by an application running in a host CPU, a notification to an application in a coprocessor/accelerator indicating that inbound data generated by the application is ready, receiving, by the application in the coprocessor/accelerator, the notification and delegating, by the application in the coprocessor/accelerator, an IO command to the application running in the host CPU, forwarding, by the application running in the host CPU, the IO command to an OS of the host CPU, transmitting, by the OS of the host CPU, an IO request to an IO device, initiating, by the IO device, a P2PDMA to transmit data associated with the IO request to a memory of the coprocessor/accelerator, and processing, by the application in the coprocessor/accelerator, the data.
1. A method, comprising:
transmitting, by an application running in a host CPU, a notification to an application in a coprocessor/accelerator indicating that inbound data generated by the application is ready;
receiving, by the application in the coprocessor/accelerator, the notification and delegating, by the application in the coprocessor/accelerator, an IO command to the application running in the host CPU;
forwarding, by the application running in the host CPU, the IO command to an OS of the host CPU;
transmitting, by the OS of the host CPU, an IO request to an IO device;
initiating, by the IO device, a P2PDMA to transmit data associated with the IO request to a memory of the coprocessor/accelerator; and
processing, by the application in the coprocessor/accelerator, the data.
2. The method as recited in claim 1 , wherein the IO command transmitted by the application in the coprocessor/accelerator is sent by way of an RPC stub.
3. The method as recited in claim 1 , wherein the IO command transmitted by the application running in the coprocessor/accelerator is received by way of a proxy in the host CPU.
4. The method as recited in claim 1 , wherein communication between an RPC stub of the coprocessor/accelerator and a proxy of the host CPU takes place in an application userspace.
5. The method as recited in claim 1 , wherein the notification is transmitted by a coordinator module of the application running in the host CPU.
6. The method as recited in claim 1 , wherein the IO device comprises one of an SSD, or a NIC.
7. The method as recited in claim 1 , wherein forwarding, by the application running in the host CPU, the IO command to an OS of the host CPU, is performed by a proxy of the application running in the host CPU.
8. The method as recited in claim 1 , wherein the operations further comprise retrieving the data from an on-chip memory of the IO device.
9. The method as recited in claim 1 , wherein after the memory of the coprocessor/accelerator receives the data from the IO device, the data is forwarded, by way of an OS of the application of the coprocessor/accelerator, to a data processor of the application of the coprocessor/accelerator.
10. The method as recited in claim 1 , wherein:
the following operations each comprise transmission of a control signal:
transmitting, by the application running in the host CPU, the notification to the application in the coprocessor/accelerator indicating that the inbound data generated by the application is ready;
receiving, by the application in the coprocessor/accelerator, the notification and delegating, by the application in the coprocessor/accelerator, the IO command to the application running in the host CPU;
forwarding, by the application running in the host CPU, the IO command to the OS of the host CPU;
transmitting, by the OS of the host CPU, the IO request to the IO device; and
the following operation comprises transmission of a data signal:
initiating, by the IO device, the P2PDMA to transmit the data associated with the IO request to the memory of the coprocessor/accelerator.
11. The method as recited in claim 1 , wherein except for the P2PDMA, no other transmission of data is performed.
12. The method as recited in claim 1 , wherein:
control operations are performed only in a control plane that includes the host CPU but does not include the coprocessor/accelerator; and
data operations are performed only in a data plane that includes the coprocessor/accelerator but does not include the host CPU.
13. A method, comprising:
storing, by an application running in a coprocessor/accelerator, processed data in a memory of the coprocessor/accelerator;
delegating, by the application running in the coprocessor/accelerator, an IO command to an application running in a host CPU;
forwarding, by the application running in the host CPU, the IO command to an OS of the host CPU;
processing, by the OS of the host CPU, the IO command;
initiating, by the OS of the host CPU, an IO request to an IO device; and
initiating, by the IO device, a P2PDMA to transmit the processed data from a memory of the coprocessor/accelerator to an on-chip memory of the IO device.
14. The method as recited in claim 13 , wherein the IO command forwarded by the application in the coprocessor/accelerator is sent by way of an RPC stub.
15. The method as recited in claim 13 , wherein the IO command forwarded by the application running in the coprocessor/accelerator is received by way of a proxy in the host CPU.
16. The method as recited in claim 13 , wherein communication between an RPC stub of the coprocessor/accelerator and a proxy of the host CPU takes place in an application userspace.
17. The method as recited in claim 13 , wherein the IO device comprises one of an SSD, ora NIC.
18. The method as recited in claim 13 , wherein except for the P2PDMA, no other transmission of data is performed.
19. The method as recited in claim 13 , wherein:
control operations are performed only in a control plane that includes the host CPU but does not include the coprocessor/accelerator; and
data operations are performed only in a data plane that includes the coprocessor/accelerator but does not include the host CPU.
20. The method as recited in claim 13 , wherein prior to storage of the processed data in the memory of the coprocessor/accelerator, the processed data is transmitted to the memory from a data processor of the coprocessor/accelerator by an OS of the coprocessor/accelerator.