TensorFusion collects comprehensive metrics for monitoring GPU infrastructure, workloads, and system performance. All metrics are stored in GreptimeDB with time-series indexing.
Measurement: tf_system_metrics
Cluster-wide system statistics and operational counters.
| Tag | Description |
|---|---|
pool | GPU pool identifier |
| Field | Type | Description |
|---|---|---|
total_workers_cnt | int64 | Total active workers |
total_nodes_cnt | int64 | Total nodes in cluster |
total_allocation_fail_cnt | int64 | Cumulative allocation failures |
total_allocation_success_cnt | int64 | Cumulative successful allocations |
total_scale_up_cnt | int64 | Cumulative scale-up events |
total_scale_down_cnt | int64 | Cumulative scale-down events |
ts | timestamp | Record timestamp |
Measurement: tf_worker_resources
Resource allocation and usage per worker.
| Tag | Description |
|---|---|
worker | Worker identifier |
workload | Associated workload |
pool | GPU pool identifier |
namespace | Kubernetes namespace |
qos | Quality of Service class |
| Field | Type | Description |
|---|---|---|
tflops_request | float64 | Requested TFLOPs |
tflops_limit | float64 | TFLOPs limit |
vram_bytes_request | float64 | Requested VRAM in bytes |
vram_bytes_limit | float64 | VRAM limit in bytes |
gpu_count | int | Number of GPUs allocated |
raw_cost | float64 | Raw compute cost |
ready | bool | Worker readiness status |
ts | timestamp | Record timestamp |
Measurement: tf_node_resources
Resource allocation and utilization per node.
| Tag | Description |
|---|---|
node | Node identifier |
pool | GPU pool identifier |
phase | Node phase/status |
| Field | Type | Description |
|---|---|---|
allocated_tflops | float64 | Allocated TFLOPs |
allocated_tflops_percent | float64 | TFLOPs utilization percentage |
allocated_vram_bytes | float64 | Allocated VRAM in bytes |
allocated_vram_percent | float64 | VRAM utilization percentage |
allocated_tflops_percent_virtual | float64 | TFLOPs vs virtual capacity |
allocated_vram_percent_virtual | float64 | VRAM vs virtual capacity |
raw_cost | float64 | Node compute cost |
gpu_count | int | Number of GPUs on node |
ts | timestamp | Record timestamp |
Measurement: tf_worker_usage
Real-time worker resource usage from hypervisor.
| Tag | Description |
|---|---|
workload | Associated workload |
worker_name | Worker identifier |
namespace | Kubernetes namespace |
pool_name | GPU pool identifier |
node_name | Host node name |
uuid | GPU UUID |
| Field | Type | Description |
|---|---|---|
compute_percentage | float64 | GPU compute utilization |
compute_tflops | float64 | Actual TFLOPs usage |
memory_percentage | float64 | VRAM utilization percentage |
memory_bytes | uint64 | VRAM usage in bytes |
ts | timestamp | Record timestamp |
Measurement: tf_gpu_usage
Detailed GPU hardware metrics from hypervisor.
| Tag | Description |
|---|---|
node | Host node name |
pool | GPU pool identifier |
uuid | GPU UUID |
| Field | Type | Description |
|---|---|---|
compute_percentage | float64 | GPU compute utilization |
memory_percentage | float64 | VRAM utilization percentage |
memory_bytes | uint64 | VRAM usage in bytes |
compute_tflops | float64 | Actual TFLOPs usage |
rx | float64 | PCIe receive KB/s |
tx | float64 | PCIe transmit KB/s |
temperature | float64 | GPU temperature (°C) |
graphics_clock_mhz | float64 | Graphics clock frequency |
sm_clock_mhz | float64 | SM clock frequency |
memory_clock_mhz | float64 | Memory clock frequency |
video_clock_mhz | float64 | Video clock frequency |
power_usage | float64 | Power consumption (W) |
nvlink_rx | float64 | NVLink receive throughput |
nvlink_tx | float64 | NVLink transmit throughput |
ts | timestamp | Record timestamp |
Measurement: tf_pool_metrics
Pool level metrics and capacity summary
| Tag | Description |
|---|---|
pool | GPU pool identifier |
phase | Pool phase/status |
| Field | Type | Description |
|---|---|---|
allocated_tflops | float64 | Total Allocated(Requested) TFLOPs |
allocated_tflops_percent | float64 | Allocated(Requested) TFLOPs percentage to total capacity |
allocated_tflops_percent_virtual | float64 | Allocated(Requested) TFLOPs percentage to virtual capacity |
allocated_vram_bytes | float64 | Total Allocated(Requested) VRAM in bytes |
allocated_vram_percent | float64 | Allocated(Requested) VRAM percentage to total capacity |
allocated_vram_percent_virtual | float64 | Allocated(Requested) VRAM percentage to virtual capacity |
limited_tflops | float64 | Total Limited TFLOPs |
limited_vram_bytes | float64 | Total Limited VRAM in bytes |
limited_tflops_percent_virtual | float64 | Limited TFLOPs percentage to virtual capacity |
limited_vram_percent_virtual | float64 | Limited VRAM percentage to virtual capacity |
gpu_count | int | Number of GPUs of the pool |
ts | timestamp | Record timestamp |
TensorFusion 文档