Computing architectures are fundamentally changing with the advent of novel technologies such as disaggregated memory (that decouples memory from compute units), as well as heterogeneous architectures that integrate different types of processors (e.g., CPUs, GPUs, AI accelerators) and memory (e.g., DRAM and non-volatile memory) within a unified system. These architectures, enabled by technologies such as RDMA and CXL, provide many benefits from dynamic resource pooling to improved utilization and latency. However, they also introduce fundamental challenges that require rethinking of existing models and development of new design principles, programming paradigms, and notions of composability.

This Dagstuhl Seminar aims to convene leading researchers and practitioners from computer architectures, distributed systems, programming languages, and database systems to develop a research agenda for disruptive memory technologies and heterogeneous architectures. The seminar aims to (1) identify key challenges in semantics and verification, (2) explore practical solutions across layers of the compute stack, and (3) foster cross-community collaboration combining both practice and theory to shape the next generation of computing platforms.

The main aim of this Dagstuhl Seminar is to work towards providing rigorous formal foundations for provably correct distributed software over disaggregated and heterogeneous architectures. With this aim in mind, we address several research questions and challenges.

1. Formal semantics and models. While RDMA and CXL each have their own industrial standards bodies that specify the behaviors of these architectures, there are gaps between the intended and actual specifications. How can we develop a full formal picture of disaggregated and heterogeneous architectures that are (a) an accurate reflection of real hardware yet (b) are usable by programmers?
2. Composable abstractions and libraries. RDMA- and CXL-based systems provide opportunities to craft new abstractions that can in turn be used to construct classical distributed objects (e.g., key-value stores and online transaction processing systems). Can we produce composable developer-friendly abstractions and APIs in a principled fashion?
3. Correct and efficient implementations. The advent of new hardware architectures raises immediate questions about efficient implementations, e.g., the optimized use of synchronizing instructions such as fence and flush or global synchronization operations. How can we minimize the synchronization overhead while providing strong correctness guarantees?
4. Applications and use cases. Our final set of discussions will revolve around applications that can best take advantage of RDMA- and CXL-based systems with different memory tiers and disaggregation. We focus on classical extensions to in-memory databases, consensus protocols, data structures (e.g., queues and ring buffers), transaction processing systems, distributed file-systems, etc. Can the designs of such applications be modularized to decouple compute, memory, and storage?

Creative Commons BY 4.0

Brijesh Dongol, Michal Friedman, Ori Lahav, and Azalea Raad

This seminar qualifies for Dagstuhl's LZI Junior Researchers program. Schloss Dagstuhl wishes to enable the participation of junior scientists with a specialisation fitting for this Dagstuhl Seminar, even if they are not on the radar of the organizers. Applications by outstanding junior scientists are possible until Friday, November 28, 2025.