ACCL: The Alveo Collective Communication Library
ACCL enables MPI-like communication between network-attached FPGAs, facilitating the development of FPGA-accelerated distributed applications. There are several advantages to utilizing ACCL instead of MPI for such applications:
FPGAs communicate directly rather than through system NICs, reducing latency and putting less pressure on the system bus.
No CPU involvement in communication, as collective orchestration is performed in the FPGA.
FPGA kernels can request communication directly, reducing CPU involvement in application orchestration, as well as eliminating kernel invocation latencies.
The figure below illustrates a typical ACCL-enabled application. The FPGA contains ACCL-specific components (the CCLO Subsystem) and optionally user-defined FPGA kernels. Either these kernels or the user application running on the CPU host, via the ACCL Driver, may initiate communication primitives, which are executed by moving data across the FPGA 100-gigabit transceivers.
FPGA-accelerated and ACCL-enabled distributed application
ACCL currently supports the following communication primitives:
Send, Receive, and Put
Broadcast
Scatter
Gather and All-Gather
Reduce, All-Reduce, and Reduce-Scatter
While ACCL does provide the above communication primitives to an application, it is important to note that ACCL is not a NIC, it cannot be utilized to perform general-purpose communication between the systems it populates. The system must be equipped with a NIC to enable CPUs in the distributed system to communicate. We refer to the FPGA network as the back-end network while the NICs connect to the front-end network. In practice, these can be the same network.