Concepts

For better reading experience, you need to understand basic HCCL concepts before using this document.

Basic HCCL Concepts

Figure 1 shows the typical communication networking of HCCL.

Figure 1 Typical communication networking

The following concepts are involved in the preceding figure:

AI server: An AI server is also called a compute node, which is an umbrella term for servers consisting of 8 or 16 Ascend NPUs.
AI cluster: An AI cluster is a system used for distributed training or inference after multiple AI servers are interconnected through switching devices.
If AI servers are connected through UB switching devices, the networking is called SuperPoD networking.
Communication member (rank): A rank is the minimum logical entity that participates in communication. Each rank is assigned a unique identifier.
Communicator: A communicator is a group of communication members that describe the communication scope. Multiple communicators can be created for a computing task, and a communication member can join multiple communicators.
Communication operator: A communicator operator completes communication tasks in a communicator. In collective communication, communication operators refer to communication operations involving all members, such as Broadcast and AllReduce.
Communication algorithm: Communication algorithms are used to implement communication operators for scenarios with different network topologies, data sizes, and hardware resources.

Terms, Acronyms, and Abbreviations

**Table 1** Terms, acronyms, and abbreviations
Term	Description
NPU	Neural Network Processing Unit. An NPU uses the data-driven parallel computing architecture and is capable of efficiently processing massive multimedia data. It is dedicated to processing a large number of computing tasks in artificial intelligence applications.
HCCL	Huawei Collective Communication Library. HCCL provides collective communication solutions for data and model parallelism in single-server multi-device and multi-server multi-device scenarios.
HCOMM	Huawei Communication, providing the basic Huawei communication library.
HCCS	Huawei Cache Coherent System. HCCS is used for high-speed interconnection between CPUs and NPUs.
HCCP	Huawei Collective Communication Adaptive Protocol. It provides cross-NPU communication capabilities and shields communication protocol differences.
TOPO	Topology. TOPO refers to the network configuration or layout formed by the connections of devices in one or more local area networks (LANs).
PCIe	Peripheral Component Interconnect Express, which is a serial peripheral expansion bus standard, typically used for peripheral expansion in a computer system.
PCIe-SW	PCIe switch, which is a switch device that complies with PCIe.
QP	Queue Pair. A QP is a core communication unit of the RDMA technology. It consists of a send queue (SQ) and a receive queue (RQ) and is used to manage data transmission tasks.
SDMA	System Direct Memory Access (DMA), which allows peripherals to directly access the system memory without the intervention of CPUs.
RDMA	Remote Direct Memory Access, which transfers data from the memory of one machine to that of another without the intervention of the operating systems of the two machines. Generally, it refers to the memory access mode across networks.
RoCE	RDMA over Converged Ethernet, which is an RDMA communication mode across Ethernets.
AIV	Vector Core in AI Core.
TS	Task Scheduler.
CCU	Collective Communication Unit.