Lecture
What If Edge Servers Were Composable? CXL, COM-HPC and CAPE
- at -
- Hall C5C5.100
- Language: English
- Type: Lecture
Lecture description
"Edge servers are usually built around a fixed combination of processors, memory and accelerators. This works when workloads are predictable, but at the edge it can lead to expensive specialization and significant overprovisioning.
What changes if these resources are no longer permanently bound to a single server? CAPE, the European Union funded research project European Open Compute Architecture for Powerful Edge, is developing an open and modular edge cloud architecture in which compute, memory and accelerators can be composed according to workload requirements. The architecture combines modular COM-HPC compute nodes, heterogeneous processors and accelerators, switched PCIe and CXL fabrics, and open server management.
This session uses the CAPE architecture to explore the engineering consequences of making edge infrastructure composable. How should compute nodes, accelerators and pooled memory be discovered and assigned? Who configures and manages the fabric? Which functions belong in platform firmware and which should be exposed through standard interfaces? And when does CXL memory pooling improve resource utilization enough to justify the additional latency and system complexity?
We will examine how CAPE approaches these questions through PCIe and CXL based resource composition, Redfish and openMPMC for hardware management, and integration with Kubernetes based edge cloud orchestration. We will also distinguish what can already be implemented using current PCIe technology from capabilities that depend on the evolving CXL ecosystem.
The session will conclude with practical design lessons for building modular and heterogeneous edge systems using open standards, including which interfaces need to remain open for multi-vendor interoperability. We will also discuss how these interfaces and open source components can be sustained beyond individual research projects, including the emerging Open Federated Edge Cloud Infrastructure Alliance, OFEIA, as a possible community and governance layer."
What changes if these resources are no longer permanently bound to a single server? CAPE, the European Union funded research project European Open Compute Architecture for Powerful Edge, is developing an open and modular edge cloud architecture in which compute, memory and accelerators can be composed according to workload requirements. The architecture combines modular COM-HPC compute nodes, heterogeneous processors and accelerators, switched PCIe and CXL fabrics, and open server management.
This session uses the CAPE architecture to explore the engineering consequences of making edge infrastructure composable. How should compute nodes, accelerators and pooled memory be discovered and assigned? Who configures and manages the fabric? Which functions belong in platform firmware and which should be exposed through standard interfaces? And when does CXL memory pooling improve resource utilization enough to justify the additional latency and system complexity?
We will examine how CAPE approaches these questions through PCIe and CXL based resource composition, Redfish and openMPMC for hardware management, and integration with Kubernetes based edge cloud orchestration. We will also distinguish what can already be implemented using current PCIe technology from capabilities that depend on the evolving CXL ecosystem.
The session will conclude with practical design lessons for building modular and heterogeneous edge systems using open standards, including which interfaces need to remain open for multi-vendor interoperability. We will also discuss how these interfaces and open source components can be sustained beyond individual research projects, including the emerging Open Federated Edge Cloud Infrastructure Alliance, OFEIA, as a possible community and governance layer."
