Workshops Archive

Mosaic: Ephemeral, Capability-Aware Scheduling Across Heterogeneous Rack-Scale Systems


Workshop: 15th International Workshop on Runtime and Operating Systems for Supercomputers (ROSS 2026)

Authors: Gourav Rattihalli, Aditya Dhakal, Kunal Sharma, and Dejan Milojicic (HPE)

Abstract: Heterogeneous HPC racks combine CPU, GPU, and FPGA resources with workloads that need different hardware and software stacks. Hardware-compatible placement alone fails when persistent workers diverge in operating-system images, libraries, drivers, or runtime versions. Separate provisioning, container, and accelerator recovery paths further fragment admission and failure handling. We present Mosaic, a Slurm-complementary lifecycle controller combining version-aware admission, capability-based dispatch, and explicit reset boundaries. Workers register capabilities and observed versions; Mosaic admits compatible workers and dispatches tasks through set inclusion and profile compatibility. It supports fast per-job userspace replacement, broader diskless PXE-to-RAM worker replacement, quarantine, and requeue. On a heterogeneous rack-scale testbed, Mosaic reaches all-workers-ready for 32 concurrent workers under one controller in about 17s. Per-job images reproduce the expected package environment in 100% of runs, versus 45.83±14.55% with periodic Ansible and 24.17±10.17% with periodic reimaging. After failure, Mosaic restores a verified configuration in about 1.5s.


Back to 15th International Workshop on Runtime and Operating Systems for Supercomputers (ROSS 2026) Archive Listing

Back to Full Workshop Archive Listing