openEuler 24.03 LTS SP4 Released with New AI, SuperPoD, Kernel, and Security Capabilities
July 1, 2026 | openEuler
OpenAtom openEuler 24.03 LTS SP4 is now available. Built on Linux kernel 6.6, the release extends openEuler 24.03 LTS with new capabilities for server, cloud, and AI workloads.
The update brings enhancements across intelligent diagnosis and tuning, UnifiedBus SuperPoD reliability and usability, NPU resource partitioning, rapid recovery for inference services, agent sandboxing, compilers, and confidential computing. It reflects contributions from 2,006 developers across 39 community member organizations.
Download openEuler 24.03 LTS SP4: openEuler download page
AI innovation
Intelligent diagnosis: New tools detect interference between containers, identify likely sources, and generate recovery recommendations. By combining an intelligent diagnosis agent with a known-issue analysis agent, openEuler can automate end-to-end diagnosis, reducing troubleshooting that previously took hours and helping teams respond faster to incidents.
Intelligent tuning: OS-level performance analysis and tuning skills are now available, together with automated Top-Down bottleneck detection. These capabilities collect key system metrics, analyze performance, and recommend optimizations.
More than 30 new AI packages: The release expands support for search, recommendation, and advertising workloads, adds compatibility with Qwen, and enables the Ascend CANN software stack. It supports deployment of the Qwen 235B model, compatibility across Ascend 910B, CANN, drivers, and the OS, and the publication of related packages, containers, and compatibility lists.
XPU Turbo for LLM inference: The optimization supports transformer-based models including DeepSeek and Qwen. It uses available CPU capacity to complement accelerator-based inference, increasing concurrency and throughput while making better use of system resources.
NPU resource partitioning: xSched is a general-purpose NPU scheduling framework for small and medium-sized models. It provides preemption, time slicing, gang scheduling, and controls for compute bandwidth and device memory capacity, together with fair and real-time scheduling policies. In colocated inference environments, it allocates single-card compute resources on demand and reduces overprovisioning. In mixed training-and-inference or online-and-offline deployments, it can rapidly preempt lower-priority tasks to protect high-priority workloads and improve NPU utilization.
Rapid fault recovery: AI inference containers can now recover from faults in seconds rather than minutes. Integration with the Kubernetes ecosystem enables second-level recovery for containerized inference services.
SuperPoD innovation
IPoURMA: IPoURMA provides a standardized way to transmit IP packets in hardware over the UnifiedBus (UB) protocol. It combines UB performance with broad Ethernet compatibility, allowing UB networks to provide standard TCP/IP socket APIs and making UB easier to adopt.
RH2D multi-tier cache direct-access acceleration: RH2D asynchronously pipelines KV cache data from a remote host directly to a local device. It is designed for AI inference and training, distributed caching, heterogeneous computing, and other workloads that frequently move large data blocks from remote nodes to GPUs. The approach is particularly suited to large values and highly concurrent, multithreaded reads, reducing serial wait time during data retrieval.
QEMU support for NPU device-to-host mapping: UB Memory is positioned between the Guest OS service driver and the host UMMU. It exposes interfaces to the shared virtual memory layer and uses VUMMU hardware emulated by QEMU, together with a mapping driver, to pass guest-host memory map and unmap operations through the virtualization stack.
Advanced UnifiedBus services: UB Service Core encapsulates low-level UB capabilities and cluster topology for UnifiedBus SuperPoDs. Applications can use SuperPoD resources much like local resources, simplifying development while preserving compatibility with existing ecosystems. A suite of cluster-level system services takes advantage of the peer-to-peer interconnect; in supported workloads, the architecture can improve application performance by 30% to 50%.
UnifiedBus reliability: A reliability plug-in blocks critical events and reports them to UBPRM, helping prevent data loss and service interruption while centralizing UB event management and supporting node fault detection and recovery. Memory-pooling fault interception and notification allow applications using shared memory to respond to failures directly. In file-sharing scenarios that borrow remote memory, page-cache placement controls help contain failures and prevent a remote page-cache fault from propagating to the local node.
SuperPoD fault localization: A new tool diagnoses latency and connectivity problems in SuperPoD KV cache and URMA communication scenarios. It can narrow down faults in under five minutes and can run either as a standalone tool or as part of an external operations platform.
Cloud-native innovation
Agent sandbox: openEuler 24.03 LTS SP4 introduces a lightweight sandbox runtime for enterprise AI agents, designed for efficient token use and end-to-end security. Hardware-software co-optimization of image snapshots, combined with remote lazy loading and tiered on-demand loading, substantially reduces cold-start time. On Kunpeng SuperPoDs, shared snapshot distribution also avoids repeated image pulls. The runtime supports secure code execution and tool use by AI agents, automated workflows, and lightweight, extensible cloud IDEs.
Kernel innovation
openEuler 24.03 LTS SP4 is based on Linux kernel 6.6 and incorporates selected capabilities from newer upstream kernels alongside innovations developed in the openEuler community. The release adds a 64 KB kernel option for Arm: Arm images can install either a 4 KB or 64 KB kernel while retaining 4 KB as the default. This preserves upstream compatibility while allowing 64 KB pages to improve performance in suitable workloads.
Programmable page cache for file systems: A programmable page-cache framework addresses model-loading I/O bottlenecks in LLM inference. Applications can customize file-system caching policies in user space according to workload characteristics, improving I/O efficiency for different models.
Cross-process zero-copy transfer: Virtual memory pages from a source process can be mapped directly into the destination process address space. The feature supports both huge and standard pages and preserves access permissions.
FUSE over io_uring: FUSE can now use io_uring to improve communication performance between user-space daemons and kernel-space drivers through /dev/fuse.
Dynamic SMT: Latency-sensitive workloads continue to run while best-effort tasks are automatically throttled and yield microarchitectural resources. Best-effort tasks can still use spare capacity on primary SMT threads and idle sibling threads, improving overall CPU utilization.
Multipath RPS enhancements: Receive Packet Steering balances load across CPU cores while taking NUMA and cluster affinity into account, reducing memory-access latency and improving network throughput.
sockmap acceleration for same-host TCP: By connecting the send and receive queues of two sockets, packets can be redirected directly to the peer socket and bypass parts of the protocol stack, shortening the forwarding path and reducing latency.
HiSock enhancements: Local acceleration, packet parsing, packet capture, address translation, and neighbor handling have been enhanced for Internet workloads, improving throughput and performance under high-concurrency, low-latency network traffic.
NetKit support: Native eBPF-based kernel traffic governance supports custom, dynamically scheduled chains of multiple eBPF programs.
sched_ext programmable scheduling: Custom scheduling algorithms can be loaded and unloaded dynamically at runtime using eBPF, without recompiling the kernel. This lowers the barrier to scheduler development and enables workload-specific optimization. If a BPF program fails, the system automatically returns to the default scheduler to avoid a system hang.
cgroup v2 enhancements: Enhancements to cgroup v2 reduce the performance cost of synchronous memory reclamation at the cgroup level. They help smooth short-lived memory spikes and can improve workload performance where applicable.
Compiler optimization
The release expands LLVM for openEuler with features that improve build efficiency, reduce debug-information bloat, and extend Triton-CPU support for FlagGems operators.
Parallel compilation: An improved ThinLTO pipeline splits large files based on the call graph, increasing parallelism and making better use of multi-core Kunpeng CPUs. This shortens build times and improves developer productivity.
DWARF utilities: Enhancements reduce the growth of debug information. Shorter builds accelerate validation of LLVM optimizations and reduce release build times.
Triton-CPU and FlagGems: Full support for FlagGems operators is added, together with affinity for AArch64 SVE and SME instructions, improving usability on AArch64 systems.
Go for openEuler: This Go distribution, based on the upstream Go project, targets cloud-native and microservice workloads. It addresses performance limitations encountered with standard Go in production environments.
BiSheng JDK: Enhancements add Kunpeng-aware instruction optimizations, improve VM startup, and increase performance for big-data workloads.
ANNC: The Accelerated Neural Network Compiler focuses on computation-graph optimization, high-performance fused operators, and efficient code generation. It integrates with mainstream open source inference frameworks and multiple hardware backends to improve CPU inference for smaller recommendation models.
Virtualization innovation
GIC overcommit optimization: Performance is improved when GICv4.1 is enabled and virtual machines frequently migrate between cores in the same affinity group.
vTimer passthrough: Direct virtual-timer delivery avoids trap-and-emulate operations during timer-interrupt injection, lowering interrupt latency and improving workload performance.
Security enhancements
CCA enhancements: Arm Confidential Compute Architecture (CCA) introduces Realms as trusted execution environments that protect the confidentiality and integrity of data and code in use. For confidential computing, openEuler supports PCIe with PCIPC protection and device passthrough into a Realm, allowing device data to travel without intermediate staging or copying and protecting the confidentiality of the device data path.
Kunpeng virtCCA enhancements: virtCCA now supports UEFI boot. Parallel hardware initialization accelerates instance startup and elastic scaling, while GPT partition tables remove the 2 TB disk limit and support cloud disks at the hundreds-of-terabytes scale for big-data and AI workloads. Standardized images simplify automated operations and batch deployment, broadening the use cases for confidential virtual machines.
Hygon security enhancements: Hardware-accelerated SM4-XTS encryption protects data at rest on local and cloud disks. SM4-GCM authenticated encryption protects the confidentiality and integrity of transaction data, logs, and message queues. Support for SM algorithms also extends to Kubernetes persistent volumes and container image layers, covering database, core financial, and cloud-native workloads.
GTA remote attestation and secure resource delivery: Global Trust Authority (GTA) remote attestation now supports Kunpeng CCA confidential computing and BMC DICE-based attestation. A new GTA Resource Broker Service first verifies the runtime environment through remote attestation, then securely provisions predefined resources, such as keys, into a trusted execution environment. The keys can be used for encryption at rest, encryption in transit, and interconnect authentication. Together, these services support upper-layer confidential computing workloads, including confidential AI inference.
The openEuler community will continue to put developers at the center as it advances open source operating systems for AI and diverse computing scenarios.