Stabilize NVIDIA GPU device mapping across loaders

audit/internal/platform/benchmark.go

@@ -679,7 +679,10 @@ func runBenchmarkInterconnect(ctx context.Context, verboseLog, runDir string, gp
 		"-g", strconv.Itoa(len(gpuIndices)),
 		"--iters", strconv.Itoa(maxInt(20, spec.NCCLSec/10)),
 	}
-	env := []string{"CUDA_VISIBLE_DEVICES=" + joinIndexList(gpuIndices)}
+	env := []string{
+		"CUDA_DEVICE_ORDER=PCI_BUS_ID",
+		"CUDA_VISIBLE_DEVICES=" + joinIndexList(gpuIndices),
+	}
 	logFunc(fmt.Sprintf("NCCL interconnect: gpus=%s", joinIndexList(gpuIndices)))
 	out, err := runSATCommandCtx(ctx, verboseLog, "nccl-all-reduce.log", cmd, env, logFunc)
 	_ = os.WriteFile(filepath.Join(runDir, "nccl-all-reduce.log"), out, 0644)

audit/internal/platform/sat.go

@@ -630,7 +630,10 @@ func nvidiaVisibleDevicesEnv(gpuIndices []int) []string {
 	if len(gpuIndices) == 0 {
 		return nil
 	}
-	return []string{"CUDA_VISIBLE_DEVICES=" + joinIndexList(gpuIndices)}
+	return []string{
+		"CUDA_DEVICE_ORDER=PCI_BUS_ID",
+		"CUDA_VISIBLE_DEVICES=" + joinIndexList(gpuIndices),
+	}
 }
 
 func runAcceptancePackCtx(ctx context.Context, baseDir, prefix string, jobs []satJob, logFunc func(string)) (string, error) {

audit/internal/platform/sat_test.go

@@ -253,11 +253,14 @@ func TestNvidiaDCGMNamedDiagCommandUsesDurationAndSelection(t *testing.T) {
 
 func TestNvidiaVisibleDevicesEnvUsesSelectedGPUs(t *testing.T) {
 	env := nvidiaVisibleDevicesEnv([]int{0, 2, 4})
-	if len(env) != 1 {
-		t.Fatalf("env len=%d want 1 (%v)", len(env), env)
+	if len(env) != 2 {
+		t.Fatalf("env len=%d want 2 (%v)", len(env), env)
 	}
-	if env[0] != "CUDA_VISIBLE_DEVICES=0,2,4" {
-		t.Fatalf("env[0]=%q want CUDA_VISIBLE_DEVICES=0,2,4", env[0])
+	if env[0] != "CUDA_DEVICE_ORDER=PCI_BUS_ID" {
+		t.Fatalf("env[0]=%q want CUDA_DEVICE_ORDER=PCI_BUS_ID", env[0])
+	}
+	if env[1] != "CUDA_VISIBLE_DEVICES=0,2,4" {
+		t.Fatalf("env[1]=%q want CUDA_VISIBLE_DEVICES=0,2,4", env[1])
 	}
 }
 

iso/overlay/usr/local/bin/bee-gpu-burn

@@ -62,6 +62,8 @@ done
 echo "loader=bee-gpu-burn"
 echo "selected_gpus=${FINAL}"
 
+export CUDA_DEVICE_ORDER="PCI_BUS_ID"
+
 TMP_DIR=$(mktemp -d)
 trap 'rm -rf "${TMP_DIR}"' EXIT INT TERM
 
@@ -78,7 +80,8 @@ for id in $(echo "${FINAL}" | tr ',' ' '); do
         fi
     fi
     echo "starting gpu ${id} size=${gpu_size_mb}MB"
-    "${WORKER}" --device "${id}" --seconds "${SECONDS}" --size-mb "${gpu_size_mb}" >"${log}" 2>&1 &
+    CUDA_VISIBLE_DEVICES="${id}" \
+        "${WORKER}" --device 0 --seconds "${SECONDS}" --size-mb "${gpu_size_mb}" >"${log}" 2>&1 &
     pid=$!
     WORKERS="${WORKERS} ${pid}:${id}:${log}"
 done

iso/overlay/usr/local/bin/bee-john-gpu-stress

@@ -152,14 +152,19 @@ done
 
 [ -n "${FINAL}" ] || { echo "no NVIDIA GPUs selected after filters" >&2; exit 1; }
 
+export CUDA_DEVICE_ORDER="PCI_BUS_ID"
+export CUDA_VISIBLE_DEVICES="${FINAL}"
+
 JOHN_DEVICES=""
+local_id=1
 for id in $(echo "${FINAL}" | tr ',' ' '); do
-    opencl_id=$((id + 1))
+    opencl_id="${local_id}"
     if [ -z "${JOHN_DEVICES}" ]; then
         JOHN_DEVICES="${opencl_id}"
     else
         JOHN_DEVICES="${JOHN_DEVICES},${opencl_id}"
     fi
+    local_id=$((local_id + 1))
 done
 
 echo "loader=john"

iso/overlay/usr/local/bin/bee-nccl-gpu-stress

@@ -70,6 +70,8 @@ echo "gpu_count=${GPU_COUNT}"
 echo "range=${MIN_BYTES}..${MAX_BYTES}"
 echo "iters=${ITERS}"
 
+export CUDA_DEVICE_ORDER="PCI_BUS_ID"
+
 deadline=$(( $(date +%s) + SECONDS ))
 round=0