Lattice: Topology-aware scheduler for HPC and AI

Scheduling

Lattice is a distributed workload scheduler built in Rust... it handles both finite batch jobs and infinite inference services. The part that really catches my eye is the use of eBPF for observability to achieve topology-aware placement on shared HPC infrastructure. Most AI training setups just throw resources at the problem... but this focuses on the actual hardware layout. It makes me think about the efficiency gains when the scheduler actually knows the physical topology... but here is the thing... does the eBPF integration allow it to react to dynamic telemetry changes in the hardware... or is the topology map treated as a static configuration during the scheduling cycle?

Source

Lattice: A topology-aware scheduler for HPC and AI workloads