Files
bee/audit/internal/platform/sat_nvidia_health.go
T

127 lines
3.1 KiB
Go

package platform
import (
"errors"
"fmt"
"strconv"
"strings"
)
func nvidiaSATStatusSeverity(status string) int {
switch strings.ToUpper(strings.TrimSpace(status)) {
case "FAILED":
return 3
case "PARTIAL", "UNSUPPORTED":
return 2
case "OK":
return 1
default:
return 0
}
}
func firstLine(s string) string {
s = strings.TrimSpace(s)
if s == "" {
return ""
}
if idx := strings.IndexByte(s, '\n'); idx >= 0 {
return strings.TrimSpace(s[:idx])
}
return s
}
func nvidiaJobNeedsHealthCheck(job satJob) bool {
if job.collectGPU {
return true
}
name := strings.ToLower(strings.TrimSpace(job.name))
return strings.Contains(name, "dcgmi") ||
strings.Contains(name, "all-reduce") ||
strings.Contains(name, "gpu-burn") ||
strings.Contains(name, "gpu-stress") ||
strings.Contains(name, "dcgmproftester")
}
func checkNvidiaJobHealth(selected []int) (string, error) {
health, _ := readNvidiaGPUHealth()
var bad []nvidiaGPUHealth
selectedSet := make(map[int]struct{}, len(selected))
for _, idx := range selected {
selectedSet[idx] = struct{}{}
}
for _, gpu := range health {
if len(selectedSet) > 0 {
if _, ok := selectedSet[gpu.Index]; !ok {
continue
}
}
if gpu.NeedsReset {
bad = append(bad, gpu)
}
}
var degraded []nvidiaNVLinkDegradedEvent
if out, err := satExecCommand("dmesg").Output(); err == nil {
for _, event := range parseNvidiaNVLinkDegradedDmesg(out) {
if event.GPUIndex >= 0 && len(selectedSet) > 0 {
if _, ok := selectedSet[event.GPUIndex]; !ok {
continue
}
}
degraded = append(degraded, event)
}
}
if len(bad) == 0 && len(degraded) == 0 {
return "", nil
}
lines := make([]string, 0, len(bad)+len(degraded)+1)
lines = append(lines, "NVIDIA GPU health check failed:")
for _, gpu := range bad {
lines = append(lines, fmt.Sprintf("gpu %d (%s) requires reset: %s", gpu.Index, gpu.Name, gpu.RawLine))
}
for _, event := range degraded {
lines = append(lines, event.Warning())
}
return strings.Join(lines, "\n"), errors.New("nvidia gpu health check failed")
}
func readNvidiaGPUHealth() ([]nvidiaGPUHealth, error) {
out, err := satExecCommand(
"nvidia-smi",
"--query-gpu=index,name,temperature.gpu,power.draw,utilization.gpu,memory.used,memory.total",
"--format=csv,noheader,nounits",
).Output()
if err != nil {
return nil, fmt.Errorf("nvidia-smi: %w", err)
}
return parseNvidiaGPUHealth(string(out)), nil
}
func parseNvidiaGPUHealth(raw string) []nvidiaGPUHealth {
var gpus []nvidiaGPUHealth
for _, line := range strings.Split(strings.TrimSpace(raw), "\n") {
line = strings.TrimSpace(line)
if line == "" {
continue
}
parts := strings.Split(line, ",")
if len(parts) < 2 {
gpus = append(gpus, nvidiaGPUHealth{RawLine: line, ParseFailure: true})
continue
}
idx, err := strconv.Atoi(strings.TrimSpace(parts[0]))
if err != nil {
gpus = append(gpus, nvidiaGPUHealth{RawLine: line, ParseFailure: true})
continue
}
upper := strings.ToUpper(line)
gpus = append(gpus, nvidiaGPUHealth{
Index: idx,
Name: strings.TrimSpace(parts[1]),
NeedsReset: strings.Contains(upper, "GPU REQUIRES RESET"),
RawLine: line,
})
}
return gpus
}