| 1 | // SPDX-License-Identifier: GPL-3.0-or-later |
| 2 | |
| 3 | package nvidia_smi |
| 4 | |
| 5 | import ( |
| 6 | "fmt" |
| 7 | "strconv" |
| 8 | "strings" |
| 9 | |
| 10 | "github.com/netdata/netdata/go/plugins/plugin/framework/collectorapi" |
| 11 | ) |
| 12 | |
| 13 | const ( |
| 14 | prioGPUPCIBandwidthUsage = collectorapi.Priority + iota |
| 15 | prioGPUPCIBandwidthUtilization |
| 16 | prioGPUFanSpeed |
| 17 | prioGPUUtilization |
| 18 | prioGPUMemUtilization |
| 19 | prioGPUDecoderUtilization |
| 20 | prioGPUEncoderUtilization |
| 21 | prioGPUMIGModeStatus |
| 22 | prioGPUMIGDevicesCount |
| 23 | prioGPUFBMemoryUsage |
| 24 | prioGPUMIGFBMemoryUsage |
| 25 | prioGPUBAR1MemoryUsage |
| 26 | prioGPUMIGBAR1MemoryUsage |
| 27 | prioGPUTemperatureChart |
| 28 | prioGPUVoltageChart |
| 29 | prioGPUClockFreq |
| 30 | prioGPUPowerDraw |
| 31 | prioGPUPerformanceState |
| 32 | ) |
| 33 | |
| 34 | var ( |
| 35 | gpuXMLCharts = collectorapi.Charts{ |
| 36 | gpuPCIBandwidthUsageChartTmpl.Copy(), |
| 37 | gpuPCIBandwidthUtilizationChartTmpl.Copy(), |
| 38 | gpuFanSpeedPercChartTmpl.Copy(), |
| 39 | gpuUtilizationChartTmpl.Copy(), |
| 40 | gpuMemUtilizationChartTmpl.Copy(), |
| 41 | gpuDecoderUtilizationChartTmpl.Copy(), |
| 42 | gpuEncoderUtilizationChartTmpl.Copy(), |
| 43 | gpuMIGModeCurrentStatusChartTmpl.Copy(), |
| 44 | gpuMIGDevicesCountChartTmpl.Copy(), |
| 45 | gpuFrameBufferMemoryUsageChartTmpl.Copy(), |
| 46 | gpuBAR1MemoryUsageChartTmpl.Copy(), |
| 47 | gpuVoltageChartTmpl.Copy(), |
| 48 | gpuTemperatureChartTmpl.Copy(), |
| 49 | gpuClockFreqChartTmpl.Copy(), |
| 50 | gpuPowerDrawChartTmpl.Copy(), |
| 51 | gpuPerformanceStateChartTmpl.Copy(), |
| 52 | } |
| 53 | migDeviceXMLCharts = collectorapi.Charts{ |
| 54 | migDeviceFrameBufferMemoryUsageChartTmpl.Copy(), |
| 55 | migDeviceBAR1MemoryUsageChartTmpl.Copy(), |
| 56 | } |
| 57 | ) |
| 58 | |
| 59 | var ( |
| 60 | gpuPCIBandwidthUsageChartTmpl = collectorapi.Chart{ |
| 61 | ID: "gpu_%s_pcie_bandwidth_usage", |
| 62 | Title: "PCI Express Bandwidth Usage", |
| 63 | Units: "B/s", |
| 64 | Fam: "pcie bandwidth", |
| 65 | Ctx: "nvidia_smi.gpu_pcie_bandwidth_usage", |
| 66 | Type: collectorapi.Area, |
| 67 | Priority: prioGPUPCIBandwidthUsage, |
| 68 | Dims: collectorapi.Dims{ |
| 69 | {ID: "gpu_%s_pcie_bandwidth_usage_rx", Name: "rx"}, |
| 70 | {ID: "gpu_%s_pcie_bandwidth_usage_tx", Name: "tx", Mul: -1}, |
| 71 | }, |
| 72 | } |
| 73 | gpuPCIBandwidthUtilizationChartTmpl = collectorapi.Chart{ |
| 74 | ID: "gpu_%s_pcie_bandwidth_utilization", |
| 75 | Title: "PCI Express Bandwidth Utilization", |
| 76 | Units: "percentage", |
| 77 | Fam: "pcie bandwidth", |
| 78 | Ctx: "nvidia_smi.gpu_pcie_bandwidth_utilization", |
| 79 | Priority: prioGPUPCIBandwidthUtilization, |
| 80 | Dims: collectorapi.Dims{ |
| 81 | {ID: "gpu_%s_pcie_bandwidth_utilization_rx", Name: "rx", Div: 100}, |
| 82 | {ID: "gpu_%s_pcie_bandwidth_utilization_tx", Name: "tx", Div: 100}, |
| 83 | }, |
| 84 | } |
| 85 | gpuFanSpeedPercChartTmpl = collectorapi.Chart{ |
| 86 | ID: "gpu_%s_fan_speed_perc", |
| 87 | Title: "Fan speed", |
| 88 | Units: "%", |
| 89 | Fam: "fan speed", |
| 90 | Ctx: "nvidia_smi.gpu_fan_speed_perc", |
| 91 | Priority: prioGPUFanSpeed, |
| 92 | Dims: collectorapi.Dims{ |
| 93 | {ID: "gpu_%s_fan_speed_perc", Name: "fan_speed"}, |
| 94 | }, |
| 95 | } |
| 96 | gpuUtilizationChartTmpl = collectorapi.Chart{ |
| 97 | ID: "gpu_%s_gpu_utilization", |
| 98 | Title: "GPU utilization", |
| 99 | Units: "%", |
| 100 | Fam: "gpu utilization", |
| 101 | Ctx: "nvidia_smi.gpu_utilization", |
| 102 | Priority: prioGPUUtilization, |
| 103 | Dims: collectorapi.Dims{ |
| 104 | {ID: "gpu_%s_gpu_utilization", Name: "gpu"}, |
| 105 | }, |
| 106 | } |
| 107 | gpuMemUtilizationChartTmpl = collectorapi.Chart{ |
| 108 | ID: "gpu_%s_memory_utilization", |
| 109 | Title: "Memory utilization", |
| 110 | Units: "%", |
| 111 | Fam: "mem utilization", |
| 112 | Ctx: "nvidia_smi.gpu_memory_utilization", |
| 113 | Priority: prioGPUMemUtilization, |
| 114 | Dims: collectorapi.Dims{ |
| 115 | {ID: "gpu_%s_mem_utilization", Name: "memory"}, |
| 116 | }, |
| 117 | } |
| 118 | gpuDecoderUtilizationChartTmpl = collectorapi.Chart{ |
| 119 | ID: "gpu_%s_decoder_utilization", |
| 120 | Title: "Decoder utilization", |
| 121 | Units: "%", |
| 122 | Fam: "dec utilization", |
| 123 | Ctx: "nvidia_smi.gpu_decoder_utilization", |
| 124 | Priority: prioGPUDecoderUtilization, |
| 125 | Dims: collectorapi.Dims{ |
| 126 | {ID: "gpu_%s_decoder_utilization", Name: "decoder"}, |
| 127 | }, |
| 128 | } |
| 129 | gpuEncoderUtilizationChartTmpl = collectorapi.Chart{ |
| 130 | ID: "gpu_%s_encoder_utilization", |
| 131 | Title: "Encoder utilization", |
| 132 | Units: "%", |
| 133 | Fam: "enc utilization", |
| 134 | Ctx: "nvidia_smi.gpu_encoder_utilization", |
| 135 | Priority: prioGPUEncoderUtilization, |
| 136 | Dims: collectorapi.Dims{ |
| 137 | {ID: "gpu_%s_encoder_utilization", Name: "encoder"}, |
| 138 | }, |
| 139 | } |
| 140 | gpuMIGModeCurrentStatusChartTmpl = collectorapi.Chart{ |
| 141 | ID: "gpu_%s_mig_mode_current_status", |
| 142 | Title: "MIG current mode", |
| 143 | Units: "status", |
| 144 | Fam: "mig", |
| 145 | Ctx: "nvidia_smi.gpu_mig_mode_current_status", |
| 146 | Priority: prioGPUMIGModeStatus, |
| 147 | Dims: collectorapi.Dims{ |
| 148 | {ID: "gpu_%s_mig_current_mode_enabled", Name: "enabled"}, |
| 149 | {ID: "gpu_%s_mig_current_mode_disabled", Name: "disabled"}, |
| 150 | }, |
| 151 | } |
| 152 | gpuMIGDevicesCountChartTmpl = collectorapi.Chart{ |
| 153 | ID: "gpu_%s_mig_devices_count", |
| 154 | Title: "MIG devices", |
| 155 | Units: "devices", |
| 156 | Fam: "mig", |
| 157 | Ctx: "nvidia_smi.gpu_mig_devices_count", |
| 158 | Priority: prioGPUMIGDevicesCount, |
| 159 | Dims: collectorapi.Dims{ |
| 160 | {ID: "gpu_%s_mig_devices_count", Name: "mig"}, |
| 161 | }, |
| 162 | } |
| 163 | gpuFrameBufferMemoryUsageChartTmpl = collectorapi.Chart{ |
| 164 | ID: "gpu_%s_frame_buffer_memory_usage", |
| 165 | Title: "Frame buffer memory usage", |
| 166 | Units: "B", |
| 167 | Fam: "fb mem usage", |
| 168 | Ctx: "nvidia_smi.gpu_frame_buffer_memory_usage", |
| 169 | Type: collectorapi.Stacked, |
| 170 | Priority: prioGPUFBMemoryUsage, |
| 171 | Dims: collectorapi.Dims{ |
| 172 | {ID: "gpu_%s_frame_buffer_memory_usage_free", Name: "free"}, |
| 173 | {ID: "gpu_%s_frame_buffer_memory_usage_used", Name: "used"}, |
| 174 | {ID: "gpu_%s_frame_buffer_memory_usage_reserved", Name: "reserved"}, |
| 175 | }, |
| 176 | } |
| 177 | gpuBAR1MemoryUsageChartTmpl = collectorapi.Chart{ |
| 178 | ID: "gpu_%s_bar1_memory_usage", |
| 179 | Title: "BAR1 memory usage", |
| 180 | Units: "B", |
| 181 | Fam: "bar1 mem usage", |
| 182 | Ctx: "nvidia_smi.gpu_bar1_memory_usage", |
| 183 | Type: collectorapi.Stacked, |
| 184 | Priority: prioGPUBAR1MemoryUsage, |
| 185 | Dims: collectorapi.Dims{ |
| 186 | {ID: "gpu_%s_bar1_memory_usage_free", Name: "free"}, |
| 187 | {ID: "gpu_%s_bar1_memory_usage_used", Name: "used"}, |
| 188 | }, |
| 189 | } |
| 190 | gpuTemperatureChartTmpl = collectorapi.Chart{ |
| 191 | ID: "gpu_%s_temperature", |
| 192 | Title: "Temperature", |
| 193 | Units: "Celsius", |
| 194 | Fam: "temperature", |
| 195 | Ctx: "nvidia_smi.gpu_temperature", |
| 196 | Priority: prioGPUTemperatureChart, |
| 197 | Dims: collectorapi.Dims{ |
| 198 | {ID: "gpu_%s_temperature", Name: "temperature"}, |
| 199 | }, |
| 200 | } |
| 201 | gpuVoltageChartTmpl = collectorapi.Chart{ |
| 202 | ID: "gpu_%s_voltage", |
| 203 | Title: "Voltage", |
| 204 | Units: "V", |
| 205 | Fam: "voltage", |
| 206 | Ctx: "nvidia_smi.gpu_voltage", |
| 207 | Priority: prioGPUVoltageChart, |
| 208 | Dims: collectorapi.Dims{ |
| 209 | {ID: "gpu_%s_voltage", Name: "voltage", Div: 1000}, // mV => V |
| 210 | }, |
| 211 | } |
| 212 | gpuClockFreqChartTmpl = collectorapi.Chart{ |
| 213 | ID: "gpu_%s_clock_freq", |
| 214 | Title: "Clock current frequency", |
| 215 | Units: "MHz", |
| 216 | Fam: "clocks", |
| 217 | Ctx: "nvidia_smi.gpu_clock_freq", |
| 218 | Priority: prioGPUClockFreq, |
| 219 | Dims: collectorapi.Dims{ |
| 220 | {ID: "gpu_%s_graphics_clock", Name: "graphics"}, |
| 221 | {ID: "gpu_%s_video_clock", Name: "video"}, |
| 222 | {ID: "gpu_%s_sm_clock", Name: "sm"}, |
| 223 | {ID: "gpu_%s_mem_clock", Name: "mem"}, |
| 224 | }, |
| 225 | } |
| 226 | gpuPowerDrawChartTmpl = collectorapi.Chart{ |
| 227 | ID: "gpu_%s_power_draw", |
| 228 | Title: "Power draw", |
| 229 | Units: "Watts", |
| 230 | Fam: "power draw", |
| 231 | Ctx: "nvidia_smi.gpu_power_draw", |
| 232 | Priority: prioGPUPowerDraw, |
| 233 | Dims: collectorapi.Dims{ |
| 234 | {ID: "gpu_%s_power_draw", Name: "power_draw"}, |
| 235 | }, |
| 236 | } |
| 237 | gpuPerformanceStateChartTmpl = collectorapi.Chart{ |
| 238 | ID: "gpu_%s_performance_state", |
| 239 | Title: "Performance state", |
| 240 | Units: "state", |
| 241 | Fam: "performance state", |
| 242 | Ctx: "nvidia_smi.gpu_performance_state", |
| 243 | Priority: prioGPUPerformanceState, |
| 244 | Dims: collectorapi.Dims{ |
| 245 | {ID: "gpu_%s_performance_state_P0", Name: "P0"}, |
| 246 | {ID: "gpu_%s_performance_state_P1", Name: "P1"}, |
| 247 | {ID: "gpu_%s_performance_state_P2", Name: "P2"}, |
| 248 | {ID: "gpu_%s_performance_state_P3", Name: "P3"}, |
| 249 | {ID: "gpu_%s_performance_state_P4", Name: "P4"}, |
| 250 | {ID: "gpu_%s_performance_state_P5", Name: "P5"}, |
| 251 | {ID: "gpu_%s_performance_state_P6", Name: "P6"}, |
| 252 | {ID: "gpu_%s_performance_state_P7", Name: "P7"}, |
| 253 | {ID: "gpu_%s_performance_state_P8", Name: "P8"}, |
| 254 | {ID: "gpu_%s_performance_state_P9", Name: "P9"}, |
| 255 | {ID: "gpu_%s_performance_state_P10", Name: "P10"}, |
| 256 | {ID: "gpu_%s_performance_state_P11", Name: "P11"}, |
| 257 | {ID: "gpu_%s_performance_state_P12", Name: "P12"}, |
| 258 | {ID: "gpu_%s_performance_state_P13", Name: "P13"}, |
| 259 | {ID: "gpu_%s_performance_state_P14", Name: "P14"}, |
| 260 | {ID: "gpu_%s_performance_state_P15", Name: "P15"}, |
| 261 | }, |
| 262 | } |
| 263 | ) |
| 264 | |
| 265 | func (c *Collector) addGpuCharts(gpu gpuInfo, index int) { |
| 266 | charts := gpuXMLCharts.Copy() |
| 267 | |
| 268 | if !isValidValue(gpu.Utilization.GpuUtil) { |
| 269 | _ = charts.Remove(gpuUtilizationChartTmpl.ID) |
| 270 | } |
| 271 | if !isValidValue(gpu.Utilization.MemoryUtil) { |
| 272 | _ = charts.Remove(gpuMemUtilizationChartTmpl.ID) |
| 273 | } |
| 274 | if !isValidValue(gpu.Utilization.DecoderUtil) { |
| 275 | _ = charts.Remove(gpuDecoderUtilizationChartTmpl.ID) |
| 276 | } |
| 277 | if !isValidValue(gpu.Utilization.EncoderUtil) { |
| 278 | _ = charts.Remove(gpuEncoderUtilizationChartTmpl.ID) |
| 279 | } |
| 280 | if !isValidValue(gpu.MIGMode.CurrentMIG) { |
| 281 | _ = charts.Remove(gpuMIGModeCurrentStatusChartTmpl.ID) |
| 282 | _ = charts.Remove(gpuMIGDevicesCountChartTmpl.ID) |
| 283 | } |
| 284 | if !isValidValue(gpu.FanSpeed) { |
| 285 | _ = charts.Remove(gpuFanSpeedPercChartTmpl.ID) |
| 286 | } |
| 287 | |
| 288 | mx := make(map[string]int64) |
| 289 | addGPUPowerMetricsSwitch(mx, "", gpu) |
| 290 | |
| 291 | if len(mx) == 0 { |
| 292 | _ = charts.Remove(gpuPowerDrawChartTmpl.ID) |
| 293 | } |
| 294 | if !isValidValue(gpu.Voltage.GraphicsVolt) { |
| 295 | _ = charts.Remove(gpuVoltageChartTmpl.ID) |
| 296 | } |
| 297 | |
| 298 | for _, c := range *charts { |
| 299 | c.ID = fmt.Sprintf(c.ID, strings.ToLower(gpu.UUID)) |
| 300 | c.Labels = []collectorapi.Label{ |
| 301 | {Key: "index", Value: strconv.Itoa(index)}, |
| 302 | {Key: "uuid", Value: gpu.UUID}, |
| 303 | {Key: "product_name", Value: gpu.ProductName}, |
| 304 | } |
| 305 | for _, d := range c.Dims { |
| 306 | d.ID = fmt.Sprintf(d.ID, gpu.UUID) |
| 307 | } |
| 308 | } |
| 309 | |
| 310 | if err := c.Charts().Add(*charts...); err != nil { |
| 311 | c.Warning(err) |
| 312 | } |
| 313 | } |
| 314 | |
| 315 | var ( |
| 316 | migDeviceFrameBufferMemoryUsageChartTmpl = collectorapi.Chart{ |
| 317 | ID: "mig_instance_%s_gpu_%s_frame_buffer_memory_usage", |
| 318 | Title: "MIG Frame buffer memory usage", |
| 319 | Units: "B", |
| 320 | Fam: "fb mem usage", |
| 321 | Ctx: "nvidia_smi.gpu_mig_frame_buffer_memory_usage", |
| 322 | Type: collectorapi.Stacked, |
| 323 | Priority: prioGPUMIGFBMemoryUsage, |
| 324 | Dims: collectorapi.Dims{ |
| 325 | {ID: "mig_instance_%s_gpu_%s_frame_buffer_memory_usage_free", Name: "free"}, |
| 326 | {ID: "mig_instance_%s_gpu_%s_frame_buffer_memory_usage_used", Name: "used"}, |
| 327 | {ID: "mig_instance_%s_gpu_%s_frame_buffer_memory_usage_reserved", Name: "reserved"}, |
| 328 | }, |
| 329 | } |
| 330 | migDeviceBAR1MemoryUsageChartTmpl = collectorapi.Chart{ |
| 331 | ID: "mig_instance_%s_gpu_%s_bar1_memory_usage", |
| 332 | Title: "MIG BAR1 memory usage", |
| 333 | Units: "B", |
| 334 | Fam: "bar1 mem usage", |
| 335 | Ctx: "nvidia_smi.gpu_mig_bar1_memory_usage", |
| 336 | Type: collectorapi.Stacked, |
| 337 | Priority: prioGPUMIGBAR1MemoryUsage, |
| 338 | Dims: collectorapi.Dims{ |
| 339 | {ID: "mig_instance_%s_gpu_%s_bar1_memory_usage_free", Name: "free"}, |
| 340 | {ID: "mig_instance_%s_gpu_%s_bar1_memory_usage_used", Name: "used"}, |
| 341 | }, |
| 342 | } |
| 343 | ) |
| 344 | |
| 345 | func (c *Collector) addMIGDeviceCharts(gpu gpuInfo, mig gpuMIGDeviceInfo) { |
| 346 | charts := migDeviceXMLCharts.Copy() |
| 347 | |
| 348 | for _, c := range *charts { |
| 349 | c.ID = fmt.Sprintf(c.ID, strings.ToLower(mig.GPUInstanceID), strings.ToLower(gpu.UUID)) |
| 350 | c.Labels = []collectorapi.Label{ |
| 351 | {Key: "gpu_uuid", Value: gpu.UUID}, |
| 352 | {Key: "gpu_product_name", Value: gpu.ProductName}, |
| 353 | {Key: "gpu_instance_id", Value: mig.GPUInstanceID}, |
| 354 | } |
| 355 | for _, d := range c.Dims { |
| 356 | d.ID = fmt.Sprintf(d.ID, mig.GPUInstanceID, gpu.UUID) |
| 357 | } |
| 358 | } |
| 359 | |
| 360 | if err := c.Charts().Add(*charts...); err != nil { |
| 361 | c.Warning(err) |
| 362 | } |
| 363 | } |
| 364 | |
| 365 | func (c *Collector) removeCharts(prefix string) { |
| 366 | prefix = strings.ToLower(prefix) |
| 367 | |
| 368 | for _, c := range *c.Charts() { |
| 369 | if strings.HasPrefix(c.ID, prefix) { |
| 370 | c.MarkRemove() |
| 371 | c.MarkNotCreated() |
| 372 | } |
| 373 | } |
| 374 | } |