LIS25-309 Compute eXpress Link, ecosystem update, future work and challenges

• Friday, 16 May 10:40 - 11:05
• Room: Session room 2 | Opala I & II

Compute eXpress Link (CXL) has evolved substantially over the last few years as the specifications and software support have matured. (http://computeexpresslink.com/) This has enabled new use cases but also brings new challenges for the software ecosystem. The main ecosystem focus remains on memory expansion and pooling solutions, but we are also seeing the first CXL attached accelerator drivers. In the memory area, elements such as improving RAS functionality, inter host sharing - including read only, and peer to peer access will keep us busy for some years to come! The origins of CXL meant that the software ecosystem initially focused on x86 support. This session will follow a brief recap of what CXL is with discussion of general ecosystem challenges and specific ‘other architecture’ issues that we encounter when using CXL on arm64 platforms. We will also touch on a few lessons that can be learnt from x86 quirks that are causing challenges for the Linux kernel. One challenge seen on arm64 is that of enabling the necessary cache flushes to allow remapping of host to device physical address spaces; a common operation for hosts using memory pools. How such flushes by Physical Address is done varies across arm64 platforms, so unlike on x86, a flexible solution is needed. We will discussion both native driver approaches and firmware mediated solutions. The final section of the session will touch on the challenges around optimal data placement that heterogenous memory brings, some approaches to which rely on arm architecture specific optimizations, and others rely on common infrastructure across diverse platforms. The aim of this session is to give the audience a heads up on where this crucial data center technology is going and a few of the problems we need to solve to get there.

Download Slides

Add to My Agenda