openEuler is an open-source operating system. The current openEuler kernel is based on Linux and supports multiple processors. It fully unleashes the potential of computing chips. As an efficient, stable, and secure open-source OS built by global open-source contributors, openEuler applies to database, big data, cloud computing, and artificial intelligence (AI) scenarios. # Key Features ## GMEM Generalized Memory Management (GMEM) is an optimal solution for memory management in OS for AI. It provides a centralized management mechanism for heterogeneous memory interconnection. GMEM innovates the memory management architecture in the Linux kernel. Its logical mapping system masks the differences between the ways how the CPU and accelerator access memory addresses. The Remote Pager memory message interaction framework provides the device access abstraction layer. In the unified address space, GMEM automatically migrates data to the OS or accelerator when data needs to be accessed or paged. GMEM APIs are consistent with native Linux memory management APIs, and feature high usability, performance, and portability. - **Logical mapping system**: GMEM high-level APIs in the kernel allow the accelerator driver to directly obtain memory management functions and create logical page tables. The logical page tables decouple the high-layer logic of memory management from the hardware layer of the CPU, so as to abstract the high-layer memory management logic that can be reused by various accelerators. - **Remote Pager**: This framework has the message channel, process management, memory swap, and memory prefetch modules for the interaction between the host and accelerator. The remote_pager abstraction layer simplifies device adaptation by enabling third-party accelerators to easily access the GMEM system. - **User APIs**: Users can directly use the memory map (mmap) of the OS to allocate the unified virtual memory. GMEM adds the flag (MMAP_PEER_SHARED) for allocating the unified virtual memory to the mmap system call. The libgmem user-mode library provides the hmadvise API of memory prefetch semantics to help users optimize the accelerator memory access efficiency. ## Native Support for Open Source Large Language Models (LLaMA and ChatGLM) The two model inference frameworks, llama.cpp and chatglm-cpp, are implemented based on C/C++. They allow users to deploy and use open source large language models on CPUs by means of model quantization. llama.cpp supports the deployment of multiple open source LLMs, such as LLaMa, LLaMa2, and Vicuna. It supports the deployment of multiple open source Chinese LLMs, such as ChatGLM-6B, ChatGLM2-6B, and Baichuan-13B. - Implemented in GGML-based C/C++. - They accelerate memory for efficient CPU inference through int4/int8 quantization, optimized KV cache, and parallel computing. ## Features in the openEuler Kernel 6.6 openEuler 24.03 LTS is built on the Linux Kernel 6.6, incorporating beneficial features and innovations from higher versions and the community. **Upstream Inherited Features:** - **Memory Management Folio Feature:** - **EEVDF Scheduling:** - **cgroup v2 Features:** - **Maple Tree and Per VMA Lock Features:** - **PCP High Adaptive Adjustment:** - **MGLRU Feature:** - **DAMON Feature:** - **Memory-Tiering Feature:** - **hugetlb vmemmap Feature:** - **huge vmalloc Feature:** - **Support for memfd_secret Syscall:** - **BIG TCP Feature:** - **XDP Multi-Buffer Support:** - **Thread-Based NAPI Polling Support:** - **BPF New kfunc Feature:** - **BPF New Dynamic Pointer Feature:** - **Perf New Feature Support:** **openEuler Key Contributions:** - **ext4 Filesystem Support for Large Folio:** - **Tidal Affinity Scheduling Feature:** - **MPAM Feature:** - **SMT Eviction Priority Reversal Feature:** - **Mixed Workload Multi-Priority:** - **Programmable Scheduling:** - **Supports arm64 vcpu hotplug:** - **Load and Computing Power Collaboration:** - **Write Access Control for Mounted Filesystem Devices:** - **RISC-V BPF new feature support:** - **ARM64 BPF new feature support:** - **Memory Reliability Classification (inherited feature):**