Communication protocol for accelerating artificial intelligence processing tasks over non-volatile memory express protocol
A method and system for data transport of artificial intelligence (AI) tasks between AI resources are provided. The method includes establishing a connection between an AI client and an AI server with a storage element capabilities; encapsulating a request to process an AI task in a command capsule compliant with a storage data transfer protocol, wherein the command capsule is encapsulated at the AI client; and transporting the command capsule to an AI server over a transport link using the storage data transfer protocol.
1. A method for data transport of artificial intelligence (AI) tasks between AI resources, comprising:
establishing a connection between an AI client and an AI server with a storage element capabilities;
encapsulating a request to process an AI task in a command capsule compliant with a storage data transfer protocol, wherein the command capsule is encapsulated at the AI client; and
transporting the command capsule to an AI server over a transport link using the storage data transfer protocol.
2. The method of claim 1 , further comprising:
encapsulating a response to the request to process the AI task in a response capsule compliant with the storage data transfer protocol, wherein the response capsule is encapsulated at the AI server; and
transporting the response capsule to the AI client over the transport link using the storage data transfer protocol, wherein a structure of the command capsule is different from a structure of the response capsule.
3. The method of claim 2 , wherein the storage data transfer protocol is a NVMeoF protocol.
4. The method of claim 3 , wherein the AI client contains a host and the AI server contains an NVMe controller.
5. The method of claim 4 , wherein the command capsule is encapsulated by the host and read by the NVMe controller.
6. The method of claim 4 , wherein the response capsule is encapsulated by the NVMe controller and sent to the host.
7. The method of claim 1 , further comprising:
forming an AI data frame including the request, wherein the AI data frame is compliant with a communication protocol that is different from the storage data transfer protocol.
8. The method of claim 7 , further comprising:
transporting the AI data frame using any one of: a read command and a write command of the storage data transfer protocol.
9. The method of claim 8 , further comprising: transporting the AI data frame by accessing specific reserved data addresses defined by the storage data transfer protocol.
10. The method of claim 7 , wherein the AI data frame includes a header portion and a payload portion, wherein the payload portion is structured to carry the data to process the AI task.
11. The method of claim 8 , further comprising:
fragmenting the AI data frame into consecutive transport protocol packets; and
encapsulating the resulting fragments of the AI data frame in payload portions of NVMeoF capsules.
12. The method of claim 1 , further comprising:
encapsulating a load computational graph (CG) in a command capsule, wherein the CG causes loading of a computational storage (CS) program; and
encapsulating an AI task request in the command capsule to cause execution of the CS program.
13. The method of claim 12 , wherein the AI task request to cause execution of the CS program defines an execution stage of the CS program, wherein the execution stage include: a discovery stage, a configuration stage, and an operation stage.
14. The method of claim 13 , wherein the CS program is any one of: load a computational graph, execute a computational graph, read from a memory, write to a memory, get memory information, and get computational graph slot information.
15. The method of claim 1 , wherein a physical medium of a network is at least a fabric.
16. A non-transitory computer readable medium having stored thereon instructions for causing a processing circuitry to execute the method of claim 1 .
17. A system for data transport of artificial intelligence (AI) tasks between AI resources, comprising, comprising:
a processing circuitry; and
a memory, the memory containing instructions that, when executed by the processing circuitry, configure the system to:
establish a connection between an AI client and an AI server with a storage element capabilities;
encapsulate a request to process an AI task in a command capsule compliant with a storage data transfer protocol, wherein the command capsule is encapsulated at the AI client; and
transport the command capsule to an AI server over a transport link using the storage data transfer protocol.