feat(presto/slurm): CPU NUMA worker placement and UCX row exchange

[misiugodfrey]

Summary

Adds runtime support for launching multiple CPU Presto workers per Slurm node with per-worker NUMA binding and UCX row exchange transport. This is the largest behavior change in the bench33 series.

Based on kjmph's bench33 POC patch.

• numa-worker-launch.sh (new): thin numactl wrapper that binds a worker process to its assigned NUMA node (cpunodebind + membind) and verifies Cpus_allowed_list / Mems_allowed_list from /proc/self/status before exec; preserves PID across the diagnostic shell so perf can attach
• prepare_cpu_numa_layout (new function in functions.sh): discovers NUMA topology across all allocated nodes via srun --overlap; builds a per-worker-slot NUMA node assignment table; rejects heterogeneous allocations rather than silently applying mismatched limits
• run_worker CPU path: spawns one srun per worker slot with computed HTTP port offsets and NUMA node; each slot runs through numa-worker-launch.sh
• UCX address resolution: resolve_node_ipv4_for_interface / resolve_node_ipv4_for_ucx_device resolve the per-GPU/per-rail internal IP used for each worker's presto.internal-address; handles multi-rail setups via CLUSTER_UCX_NET_DEVICES_BY_LOCAL_WORKER
• --verify-cpu-ucx: pre-benchmark check that all expected worker UCX ports are responding before query execution starts
• launch-run.sh: new flags --cpu, --use-numa, --workers-per-host N, --ucx-net-device, --internal-address-interface, --verify-cpu-ucx
• worker.env: baseline CPU UCX runtime knobs: VELOX_UCX_CPU_EXCHANGE, PRESTO_WORKER_UCX_LOCAL_LIB_FIRST, UCX_MODULE_DIR, UCX_TLS=sysv,tcp,self, UCX_NET_DEVICES, UCX_POSIX_USE_PROC_LINK=n
• launch-analyze-tables.sh: CPU variant support, --worker-env-file option

Dependencies

• Depends on PR More GHA experiments #2 (bench-numa-config) for the NUMA topology discovery helpers in common_functions.sh
• Depends on PR Land @bdice prototype workflows in main for testing #3 (bench-slurm-robustness) for srun PID tracking and result dir structure
• Rebase on PR Land @bdice prototype workflows in main for testing #3 after it merges to drop duplicated hunks in the split files

Verification

• On a 2-socket Grace host with CPU_WORKERS_PER_HOST=2, launch a CPU benchmark with --use-numa and verify each worker's container is bound to a distinct NUMA node (check Cpus_allowed_list in worker logs)
• Verify per-worker HTTP port offsets are unique (e.g. workers at ports 9080, 9083, 9086 for a 3-worker-per-host setup)
• Run with --verify-cpu-ucx and confirm the preflight check blocks until all UCX ports are responding
• Run launch-analyze-tables.sh --cpu and verify it successfully runs the analyze job with CPU workers
• Verify worker.env UCX knobs are picked up: UCX_TLS=sysv,tcp,self and UCX_POSIX_USE_PROC_LINK=n should appear in worker environment
• On a multi-rail setup with CLUSTER_UCX_NET_DEVICES_BY_LOCAL_WORKER, verify each worker's UCX_NET_DEVICES is set to the correct per-slot NIC
• Run ./ci/check_style.sh on changed files

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.