Open Source Summit North America 2025

Graphics Processing Units (GPUs) have become key to modern computing, enabling high-performance parallel processing beyond their original role in gaming. Today, they are critical for solving complex computational challenges effectively. To achieve high levels of parallelism, OpenCL is commonly used to break tasks into kernels, which are executed by multiple threads. Optimizing OpenCL performance on GPUs remains a persistent challenge, as it involves fine-tuning both software (OpenCL code, kernels) and hardware. One of the critical factors influencing GPU performance is the effective use of workgroups. The size of workgroups significantly impacts parallelism and efficiency.
This session focuses on key software optimizations in OpenCL, such as efficient memory management, kernel fusion and optimal workgroup sizing in improving GPU performance. We will demonstrate, how optimized OpenCL code can significantly enhance parallel execution and efficiency. This talk also addresses challenges faced by automotive GPUs, including power and thermal constraints with strategies to overcome these. Best practices for writing efficient OpenCL code tailored for automotive GPUs will also be outlined.

This presentation explores how to create a multi-threaded version of Valkey by employing userspace Read-Copy-Update (RCU) to achieve high performance. With fewer than 2,000 lines of code modifications, we introduce a coordinator-worker pattern, enabling key-value stores like Valkey and Redis to handle tasks concurrently across multiple threads while using a per-thread event loop for I/O operations. Userspace RCU facilitates lock-free data sharing between a writer thread and multiple reader threads, dramatically improving read performance.

Our experiments show that the multi-threaded Valkey can achieve over one million operations per second on a standard server.

For long time kexec was a faster way to reboot a machine without incurring delays caused by firmware and bootloaders. However for many applications even a kexec reboot still means significant service degradation, like disruption of the running guests in virtualized environments and the need to rebuild in-memory caches for large databases.

We propose Kexec HandOver (KHO) mechanism that allows serialization and deserialization of kernel data as well as preserving arbitrary memory ranges across kexec.

In addition, KHO keeps physically contiguous memory regions that are guaranteed to not have any memory that KHO would preserve, but still can be used by the system. The kexeced kernel bootstraps itself using those regions and marks all handed over memory as in use. KHO users then can recover their state from the preserved data. This includes memory reservations, where the user can either discard or claim reservations.

As eBPF continues to revolutionize Linux observability and networking, the complexity of its program loading mechanism has evolved significantly.

This technical deep dive unravels the sophisticated machinery behind eBPF program loading, exploring the intricate interplay between userspace loader and Linux kernel verifier. We'll dissect the eBPF program relocation mechanisms, examine the role of BTF (BPF Type Format) in enabling strong typing and verification capabilities, and analyze the complex choreography of bpf() syscalls that bridge userspace and kernel operations. Finally, we will also discuss the security implications and program signing challenges in the loading pipeline.