epa-orchestrator

EPA Orchestrator is designed to provide secure, policy-driven resource orchestration for snaps and workloads on Linux systems. Its vision is to enable fine-grained, dynamic allocation and management of system resources—starting with CPU pinning and memory management, with plans to expand to other resource types and orchestration policies. The orchestrator exposes a secure Unix socket API for resource allocation and introspection, making it easy for other snaps (such as openstack-hypervisor) and workloads to request and manage dedicated or shared resources in a controlled manner.

Features:

• CPU Pinning and Allocation: Allocate isolated and shared CPU sets to snaps and workloads, supporting both dedicated and shared CPU usage models with basic system-size heuristics.
• Memory Management and Hugepage Tracking: Introspect NUMA hugepages and track hugepage allocations across NUMA nodes with per-service allocation tracking.
• NUMA-Aware Core Allocation: Request a specific number of cores from a particular NUMA node with override/append semantics and exact-count guarantees.
• Resource Introspection: Query current allocations and available resources via a secure API.
• Secure Unix Socket API: All orchestration actions are performed via a secure, local Unix socket with JSON-based requests and responses.
• Basic Allocation Heuristics: Automatic allocation based on system size (small vs large systems) when no specific core count is requested.

CPU Allocation Policy: Small vs. Large Systems:

When a client requests core allocation with num_of_cores: 0, EPA Orchestrator applies a policy based on the total number of CPUs detected:

• Small systems (≤100 CPUs):
  • By default, 80% of the available CPUs are allocated to the requesting snap or workload.
  • The remaining 20% are left unallocated (shared).
• Large systems (>100 CPUs):
  • By default, 16 CPUs are always reserved (left unallocated/shared).
  • All other CPUs are allocated to the requesting snap or workload.

This policy ensures that on large servers, a fixed number of CPUs are always available for system or shared use, while on smaller systems, a proportional allocation is used.

NUMA-Aware Core Allocation Policy:

The NUMA-aware allocation action allows services to request a specific number of cores from a particular NUMA node:

• NUMA Locality: Cores are allocated from the specified NUMA node to ensure optimal memory access patterns.
• Force Reallocation: NUMA allocation will override any existing non-explicit allocations to other services, even if they span multiple NUMA nodes.
• Atomic Exact-Count: If fewer than the requested number of cores are available in the NUMA node, the request fails with an error; no partial allocation occurs.
• Priority System: NUMA allocations take precedence over automatic allocations and cannot be overridden by other services.
• Per-NUMA override/append semantics: If the same service requests the same NUMA node again, it overrides previous cores from that node. If it requests a different NUMA node, the new cores are appended so the service may hold allocations across multiple NUMA nodes.
• Per-NUMA deallocation: Sending num_of_cores = -1 for a node deallocates any existing cores for that service in that node. num_of_cores = 0 is invalid for NUMA.