Disaggregated Heterogeneous Architectures

This BoF will discuss the most recent research topics in disaggregated heterogeneous architectures and identify their key drivers. It aims at establishing a long-term forum for exchange of ideas and experience. Today’s HPC systems are highly heterogeneous, tightly integrated machines combining different processors, network, memory, and storage technologies. This diversification increases with the integration of disruptive technologies, e.g. AI accelerators, neuromorphic devices, or even quantum computers. Orchestrating and using this “zoo” of very different hardware in an efficient way poses significant challenges. “Disaggregated” systems organise heterogeneous resources in “pools” of many like resources and dynamically compose these according to the application needs. This approach is established in Cloud Computing (mainly for memory), and gains traction in HPC with the “Modular Supercomputing Architecture”. Examples are Perlmutter at LBNL/NERSC, Wisteria/BDEC-01 at University of Tokyo, or JUWELS at JSC. Advances in interconnects and protocols, such as Compute Express Link (CXL), open new opportunities – for example memory pooling at the rack level. We will report on approaches and early results targeting improvements in throughput and energy efficiency of large systems by composing an optimal set of resources for each application or workflow from compute, accelerator and memory pools. This avoids over-provisioning resources to cover the needs of the most demanding applications, which is common in statically configured supercomputers, and thus saves capital expenses and energy. It does, however, require advanced scheduling, orchestration and resource management techniques to minimise configuration overheads and achieve a good match of resources to applications.

---

Contributors:

• Kengo Nakajima
• Anshu Dubey
• Nick Wright