C2MPI: a hardware-agnostic message passing interface for heterogeneous computing systems
Compute-centric message passing interface (C 2 MPI) provides a hardware-agnostic message passing interface for heterogeneous computing systems. Hardware-agnostic programming with high performance portability is envisioned to be a bedrock for realizing adoption of emerging accelerator technologies in heterogeneous computing systems, such as high-performance computing (HPC) systems, data center computing systems, and edge computing systems. The adoption of emerging accelerators is key to achieving greater scale and performance in heterogeneous computing systems. Accordingly, embodiments described herein provide a flexible hardware-agnostic environment that allows application developers to develop high-performance applications without knowledge of the underlying hardware.
1 . A method for providing instructions for a host application to a heterogeneous computing system via a compute-centric message passing interface (C 2 MPI), the method comprising:
providing a first hardware-agnostic instruction to invoke a first child rank corresponding to a first accelerator resource, wherein the first hardware-agnostic instruction specifies a first computational function,
wherein the C 2 MPI includes application parent ranks defined inside a hardware-agnostic region of the host application, and accelerator parent ranks, the application parent ranks and the accelerator parent ranks managing child ranks, the accelerator parent ranks performing hardware management, kernel retrieval, registration, and execution and maintaining of system resources allocated by the application parent ranks, and
wherein the application parent ranks and the accelerator parent ranks allocate and deallocate the child ranks based on a duration of execution time of the host application or the first accelerator resource, one of the child ranks being associated with the first accelerator resource after allocation.
2 . The method of claim 1 , further comprising providing a second hardware-agnostic instruction to send first data to the first child rank for processing.
3 . The method of claim 2 , wherein the second hardware-agnostic instruction is sent in response to receiving the first child rank.
4 . The method of claim 2 , further comprising providing a third hardware-agnostic instruction to receive a first processing result from the first child rank.
5 . The method of claim 1 , further comprising providing a second hardware-agnostic instruction to invoke a second child rank corresponding to a second accelerator resource.
6 . The method of claim 5 , wherein the second hardware-agnostic instruction specifies the first computational function.
7 . The method of claim 5 , wherein the second hardware-agnostic instruction specifies a second computational function different from the first computational function.
8 . The method of claim 5 , further comprising:
providing a third hardware-agnostic instruction to send first data to the first child rank for processing; and
providing a fourth hardware-agnostic instruction to send second data to the second child rank for processing.
9 . The method of claim 8 , wherein the third hardware-agnostic instruction and the fourth hardware agnostic instruction allow for parallel execution of the first child rank and the second child rank.
10 . The method of claim 8 , wherein the third hardware-agnostic instruction and the fourth hardware agnostic instruction require pipeline execution of the first child rank and the second child rank.
11 . A method for executing instructions for an application on a heterogeneous computing system received via a compute-centric message passing interface (C 2 MPI), the method comprising:
receiving, from a host application, a first hardware-agnostic instruction to invoke a first child rank corresponding to a first accelerator resource, wherein the first hardware-agnostic instruction specifies a first computational function; and
locating the first accelerator resource based on the first computational function,
wherein the C 2 MPI includes application parent ranks, inside a hardware-agnostic region of the application, and accelerator parent ranks, the application parent ranks and the accelerator parent ranks managing child ranks, the accelerator parent ranks performing hardware management, kernel retrieval, registration, and execution and maintaining of system resources allocated by the application parent ranks, and
wherein the application parent ranks and the accelerator parent ranks allocate and deallocate the child ranks based on a duration of execution time of the application or the first accelerator resource, one of the child ranks being associated with the first accelerator resource after allocation.
12 . The method of claim 11 , wherein the first accelerator resource corresponds to one or a set of accelerators programmed to execute the first computational function.
13 . The method of claim 12 , wherein the set of accelerators programmed to execute the first computational function comprise a same type of accelerator.
14 . The method of claim 12 , wherein the set of accelerators programmed to execute the first computational function comprise heterogeneous types of accelerators.
15 . The method of claim 11 , further comprising:
invoking the first child rank using the first accelerator resource; and returning the first child rank to the host application.
16 . The method of claim 11 , wherein locating the first accelerator resource comprises selecting a type of accelerator suitable for executing the first computational function.
17 . The method of claim 16 , wherein selecting the type of accelerator suitable for executing the first computational function comprises selecting the type of accelerator having a pre-selected performance for executing the first computational function.
18 . The method of claim 16 , wherein locating the first accelerator resource further comprises:
identifying an available one of the type of accelerator suitable for executing the first computational function; and
reserving the available one of the type of accelerator suitable for executing the first computational function as the first accelerator resource.
19 . The method of claim 11 , further comprising:
receiving a second hardware-agnostic instruction to send first data to the first child rank for processing; and
forwarding the first data to the first accelerator resource.
20 . The method of claim 19 , further comprising receiving a first processing result from the first accelerator resource corresponding to the first data.
21 . The method of claim 20 , further comprising staging the first processing result.
22 . The method of claim 20 , further comprising:
receiving a third hardware-agnostic instruction to receive the first processing result from the child rank; and
forwarding the first processing result to the host application.
23 . A non-transitory computer-readable medium having stored thereon software instructions that, when executed by a processor, cause the processor to:
receive, from a host application, a first hardware-agnostic instruction to invoke a first child rank corresponding to a first accelerator resource, wherein the first hardware-agnostic instruction specifies a first computational function;
locate the first accelerator resource based on the first computational function;
invoke the first child rank; and
return the first child rank to the host application,
wherein the C 2 MPI includes application parent ranks, inside a hardware-agnostic region of the application, and accelerator parent ranks, the application parent ranks and the accelerator parent ranks managing child ranks, the accelerator parent ranks performing hardware management, kernel retrieval, registration, and execution and maintaining of system resources allocated by the application parent ranks, and
wherein the application parent ranks and the accelerator parent ranks allocate and deallocate the child ranks based on a duration of execution time of the application or the first accelerator resource, one of the child ranks being associated with the first accelerator resource after allocation.