Overview
Learn about peer provider composability in libfabric through a comprehensive technical talk series featuring three interconnected presentations from Intel Corporation experts. Explore the architecture and API design of libfabric's peer provider system, which enables applications to leverage multiple network technologies simultaneously for optimal performance. Dive into practical implementations, including how shared memory providers pair with scale-out providers and how focused providers integrate with core providers for collective operations. Master the intricacies of the peer APIs, shared completion queues, shared receive contexts, and collective offloading mechanisms. Understand how this framework allows independent development of specialized providers that can work together seamlessly, supporting complex scenarios involving local node accelerations, GPU fabrics, HPC NICs, and various network transport configurations. Gain insights into the future of network communication as systems become increasingly heterogeneous and require sophisticated solutions for maximum performance.
Syllabus
Intro
Example: Shared Memory + Network Provider
Complex, Theoretical Scenarios
SHARED MEMORY ACCELERATION
PEER PROVIDER EXAMPLES
SHARED COMPLETION QUEUE API
SHARED RECEIVE CONTEXT
EXAMPLE SRX FLOW
OFI COLLECTIVE API
IMPLEMENTATION CONSIDERATIONS
COLLECTIVE OFFLOAD WITH PEER PROVIDER
DESIGN OVERVIEW
COLLECTIVE GROUP CREATION
JOIN COLLECTIVE GROUP
COLLECTIVE OPS
BOOTSTRAP COLLECTIVE
UTILITY COLLECTIVE PROVIDER
CONCLUSION AND FUTURE WORK
Taught by
OpenFabrics Alliance