--addr | 127.0.0.1:50055 | Bind address |
--devices | auto-detect all | CUDA device IDs, comma-separated |
--pool-size | 30gb | Pinned memory pool size |
--hint-value-size | — | Hint for typical value size to tune cache/allocator |
--use-hugepages | false | Use huge pages (requires /proc/sys/vm/nr_hugepages) |
--enable-lfu-admission | false | Enable TinyLFU admission (default: plain LRU) |
--disable-numa-affinity | false | Disable NUMA-aware allocation |
--http-addr | 0.0.0.0:9091 | HTTP server for health check and Prometheus |
--enable-prometheus | true | Enable /metrics endpoint |
--metrics-otel-endpoint | — | OTLP metrics export endpoint |
--metrics-period-secs | 5 | Metrics export period (OTLP only) |
--log-level | info | trace/debug/info/warn/error |
--ssd-cache-path | — | Enable SSD cache with file path |
--ssd-cache-capacity | 512gb | SSD cache capacity |
--ssd-write-queue-depth | 8 | Max pending write batches |
--ssd-prefetch-queue-depth | 2 | Max pending prefetch batches |
--ssd-write-inflight | 2 | Max concurrent block writes |
--ssd-prefetch-inflight | 16 | Max concurrent block reads |
--max-prefetch-blocks | 800 | Backpressure for SSD prefetch |
--trace-sample-rate | 1.0 | Sampling rate 0.0–1.0 (requires --features tracing) |
--metaserver-addr | — | MetaServer gRPC address for cross-node discovery (requires --addr to be routable) |