Open Source Summit Europe 2024

his presentation will focus on CXL (Compute Express Link) as an advanced interconnect between machines and peripherals. CXL allows to leverage the PCIe physical interconnect to link together different device types (CPU, memory, I/O, cache, switches etc) into a combined hierarchy. This allows IHVs to create tailored solutions for eg large-scale AI systems or dynamic resource pooling between machines. As it's also possible to connect or pool memory resources it means the we can end up with some really interesting NUMA topologies. Plus we need to look at memory placement, as CXL memory is inherently hotpluggable, and as such not really suitable for some data structures like DMA areas etc. In this talk I will give an overview over CXL and the implications for NUMA topologies, and I'll be giving a short demo with an emulated CXL instance under qemu.