Skip to content

fix(scheduler): graceful handling of asymmetric GPU topology to prevent panics #2217

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
Norway-02 wants to merge 3 commits into from
Closed

fix(scheduler): graceful handling of asymmetric GPU topology to prevent panics #2217

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
cecd3aa
fix(rm): resolve go vet sync.Mutex copy-by-value errors
Norway-02 Jul 28, 2026
a01f782
fix: use proto.Merge to safely copy embedded protobuf and extend tests
Norway-02 Jul 29, 2026
68b73bf
fix(scheduler): graceful handling of asymmetric GPU topology to preve…
Norway-02 Jul 30, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion pkg/device-plugin/nvidiadevice/nvinternal/plugin/server_test.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -250,7 +250,7 @@ func TestCDIAllocateResponse(t *testing.T) {
}

for i := range testCases {
tc := testCases[i]
tc := &testCases[i]
t.Run(tc.description, func(t *testing.T) {
deviceListStrategies, _ := v1.NewDeviceListStrategies(tc.deviceListStrategies)
plugin := NvidiaDevicePlugin{
Expand Down
15 changes: 12 additions & 3 deletions pkg/device-plugin/nvidiadevice/nvinternal/rm/device_map.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -41,6 +41,8 @@ import (
"k8s.io/klog/v2"

spec "github.com/NVIDIA/k8s-device-plugin/api/config/v1"
"google.golang.org/protobuf/proto"

)

type deviceMapBuilder struct {
Expand Down Expand Up @@ -340,10 +342,17 @@ func updateDeviceMapWithReplicas(replicatedResources *spec.ReplicatedResources,
for _, id := range ids {
for i := 0; i < r.Replicas; i++ {
annotatedID := string(NewAnnotatedID(id, i))
replicatedDevice := *(oDevices[r.Name][id])
orig := oDevices[r.Name][id]
replicatedDevice := &Device{
Paths: orig.Paths,
Index: orig.Index,
TotalMemory: orig.TotalMemory,
ComputeCapability: orig.ComputeCapability,
Replicas: r.Replicas,
}
proto.Merge(&replicatedDevice.Device, &orig.Device)
replicatedDevice.ID = annotatedID
replicatedDevice.Replicas = r.Replicas
devices.insert(name, &replicatedDevice)
devices.insert(name, replicatedDevice)
Comment thread
Norway-02 marked this conversation as resolved.
}
}
}
Expand Down
89 changes: 80 additions & 9 deletions pkg/device-plugin/nvidiadevice/nvinternal/rm/device_map_test.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,6 +40,7 @@ import (

spec "github.com/NVIDIA/k8s-device-plugin/api/config/v1"
"github.com/stretchr/testify/require"
"google.golang.org/protobuf/proto"
kubeletdevicepluginv1beta1 "k8s.io/kubelet/pkg/apis/deviceplugin/v1beta1"
)

Expand Down Expand Up @@ -127,8 +128,22 @@ func TestDeviceMapInsert(t *testing.T) {
}

func TestUpdateDeviceMapWithReplicas(t *testing.T) {
device0 := Device{Device: kubeletdevicepluginv1beta1.Device{ID: "0"}, Index: "0"}
device1 := Device{Device: kubeletdevicepluginv1beta1.Device{ID: "1"}}
device0 := Device{
Device: kubeletdevicepluginv1beta1.Device{ID: "0", Health: "Healthy", Topology: &kubeletdevicepluginv1beta1.TopologyInfo{Nodes: []*kubeletdevicepluginv1beta1.NUMANode{{ID: 0}}}},
Index: "0",
Paths: []string{"/dev/nvidia0"},
TotalMemory: 1024,
ComputeCapability: "8.0",
Replicas: 1,
}
device1 := Device{
Device: kubeletdevicepluginv1beta1.Device{ID: "1", Health: "Unhealthy", Topology: &kubeletdevicepluginv1beta1.TopologyInfo{Nodes: []*kubeletdevicepluginv1beta1.NUMANode{{ID: 1}}}},
Index: "1",
Paths: []string{"/dev/nvidia1"},
TotalMemory: 2048,
ComputeCapability: "8.6",
Replicas: 1,
}
device2 := Device{Device: kubeletdevicepluginv1beta1.Device{ID: "2"}}
device3 := Device{Device: kubeletdevicepluginv1beta1.Device{ID: "3"}}

Expand Down Expand Up @@ -179,10 +194,38 @@ func TestUpdateDeviceMapWithReplicas(t *testing.T) {
},
expectedDeviceMap: DeviceMap{
"replicated-resource1": Devices{
"0::0": &Device{Device: kubeletdevicepluginv1beta1.Device{ID: "0::0"}, Index: "0", Replicas: 2},
"0::1": &Device{Device: kubeletdevicepluginv1beta1.Device{ID: "0::1"}, Index: "0", Replicas: 2},
"1::0": &Device{Device: kubeletdevicepluginv1beta1.Device{ID: "1::0"}, Replicas: 2},
"1::1": &Device{Device: kubeletdevicepluginv1beta1.Device{ID: "1::1"}, Replicas: 2},
"0::0": &Device{
Device: kubeletdevicepluginv1beta1.Device{ID: "0::0", Health: "Healthy", Topology: &kubeletdevicepluginv1beta1.TopologyInfo{Nodes: []*kubeletdevicepluginv1beta1.NUMANode{{ID: 0}}}},
Index: "0",
Paths: []string{"/dev/nvidia0"},
TotalMemory: 1024,
ComputeCapability: "8.0",
Replicas: 2,
},
"0::1": &Device{
Device: kubeletdevicepluginv1beta1.Device{ID: "0::1", Health: "Healthy", Topology: &kubeletdevicepluginv1beta1.TopologyInfo{Nodes: []*kubeletdevicepluginv1beta1.NUMANode{{ID: 0}}}},
Index: "0",
Paths: []string{"/dev/nvidia0"},
TotalMemory: 1024,
ComputeCapability: "8.0",
Replicas: 2,
},
"1::0": &Device{
Device: kubeletdevicepluginv1beta1.Device{ID: "1::0", Health: "Unhealthy", Topology: &kubeletdevicepluginv1beta1.TopologyInfo{Nodes: []*kubeletdevicepluginv1beta1.NUMANode{{ID: 1}}}},
Index: "1",
Paths: []string{"/dev/nvidia1"},
TotalMemory: 2048,
ComputeCapability: "8.6",
Replicas: 2,
},
"1::1": &Device{
Device: kubeletdevicepluginv1beta1.Device{ID: "1::1", Health: "Unhealthy", Topology: &kubeletdevicepluginv1beta1.TopologyInfo{Nodes: []*kubeletdevicepluginv1beta1.NUMANode{{ID: 1}}}},
Index: "1",
Paths: []string{"/dev/nvidia1"},
TotalMemory: 2048,
ComputeCapability: "8.6",
Replicas: 2,
},
},
"resource2": Devices{
"2::0": &Device{Device: kubeletdevicepluginv1beta1.Device{ID: "2::0"}, Replicas: 1},
Expand Down Expand Up @@ -220,8 +263,22 @@ func TestUpdateDeviceMapWithReplicas(t *testing.T) {
},
expectedDeviceMap: DeviceMap{
"replicated-resource1": Devices{
"0::0": &Device{Device: kubeletdevicepluginv1beta1.Device{ID: "0::0"}, Index: "0", Replicas: 2},
"0::1": &Device{Device: kubeletdevicepluginv1beta1.Device{ID: "0::1"}, Index: "0", Replicas: 2},
"0::0": &Device{
Device: kubeletdevicepluginv1beta1.Device{ID: "0::0", Health: "Healthy", Topology: &kubeletdevicepluginv1beta1.TopologyInfo{Nodes: []*kubeletdevicepluginv1beta1.NUMANode{{ID: 0}}}},
Index: "0",
Paths: []string{"/dev/nvidia0"},
TotalMemory: 1024,
ComputeCapability: "8.0",
Replicas: 2,
},
"0::1": &Device{
Device: kubeletdevicepluginv1beta1.Device{ID: "0::1", Health: "Healthy", Topology: &kubeletdevicepluginv1beta1.TopologyInfo{Nodes: []*kubeletdevicepluginv1beta1.NUMANode{{ID: 0}}}},
Index: "0",
Paths: []string{"/dev/nvidia0"},
TotalMemory: 1024,
ComputeCapability: "8.0",
Replicas: 2,
},
},
"resource1": Devices{
"1": &device1,
Expand All @@ -233,7 +290,21 @@ func TestUpdateDeviceMapWithReplicas(t *testing.T) {
for _, tc := range testCases {
t.Run(tc.description, func(t *testing.T) {
devices, _ := updateDeviceMapWithReplicas(&tc.config.Config.Sharing.TimeSlicing, tc.devices)
require.EqualValues(t, tc.expectedDeviceMap, devices)
require.Equal(t, len(tc.expectedDeviceMap), len(devices))
for resourceName, expectedDevices := range tc.expectedDeviceMap {
actualDevices := devices[resourceName]
require.Equal(t, len(expectedDevices), len(actualDevices))
for id, expectedDevice := range expectedDevices {
actualDevice := actualDevices[id]
require.NotNil(t, actualDevice)
require.Equal(t, expectedDevice.Index, actualDevice.Index)
require.Equal(t, expectedDevice.Paths, actualDevice.Paths)
require.Equal(t, expectedDevice.TotalMemory, actualDevice.TotalMemory)
require.Equal(t, expectedDevice.ComputeCapability, actualDevice.ComputeCapability)
require.Equal(t, expectedDevice.Replicas, actualDevice.Replicas)
require.True(t, proto.Equal(&expectedDevice.Device, &actualDevice.Device), "proto fields should match")
}
}
})
}
}
Expand Down
6 changes: 3 additions & 3 deletions pkg/device-plugin/nvidiadevice/nvinternal/rm/devices.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -253,7 +253,7 @@ func (ds Devices) AlignedAllocationSupported() bool {
}

// AlignedAllocationSupported checks whether the device supports an aligned allocation
func (d Device) AlignedAllocationSupported() bool {
func (d *Device) AlignedAllocationSupported() bool {
if d.IsMigDevice() {
return false
}
Expand All @@ -268,12 +268,12 @@ func (d Device) AlignedAllocationSupported() bool {
}

// IsMigDevice returns checks whether d is a MIG device or not.
func (d Device) IsMigDevice() bool {
func (d *Device) IsMigDevice() bool {
return strings.Contains(d.Index, ":")
}

// GetUUID returns the UUID for the device from the annotated ID.
func (d Device) GetUUID() string {
func (d *Device) GetUUID() string {
return AnnotatedID(d.ID).GetID()
}

Expand Down
15 changes: 11 additions & 4 deletions pkg/device/nvidia/calculate_score.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,6 +22,7 @@ import (

"github.com/NVIDIA/go-nvlib/pkg/nvlib/device"
"github.com/NVIDIA/go-nvml/pkg/nvml"
"k8s.io/klog/v2"
)

// Device represents a GPU device as reported by NVML, including all of its
Expand Down Expand Up @@ -210,14 +211,20 @@ func calculateGPUPairScore(gpu0 *Device, gpu1 *Device) int {
return 0
}

if len(gpu0.Links[gpu1.Index]) != len(gpu1.Links[gpu0.Index]) {
err := fmt.Errorf("internal error in bestEffort GPU allocator: all P2PLinks between 2 GPUs should be bidirectional")
panic(err)
links0 := gpu0.Links[gpu1.Index]
links1 := gpu1.Links[gpu0.Index]

if len(links0) != len(links1) {
klog.Warningf("Asymmetric GPU topology detected: %s->%s=%d links, %s->%s=%d links. This may be caused by hardware issues, driver bugs, or inconsistent reporting. Evaluating common subset of links.",
gpu0.UUID, gpu1.UUID, len(links0),
gpu1.UUID, gpu0.UUID, len(links1))
}

score := 0
common := min(len(links0), len(links1))

for _, link := range gpu0.Links[gpu1.Index] {
for i := 0; i < common; i++ {
link := links0[i]
Comment thread
Norway-02 marked this conversation as resolved.
switch link.Type {
case P2PLinkCrossCPU:
score += 10
Expand Down
99 changes: 99 additions & 0 deletions pkg/device/nvidia/calculate_score_test.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -108,6 +108,105 @@ func Test_calculateGPUScore(t *testing.T) {
},
},
},
{
name: "asymmetric nvlink test gracefully handles shorter slice",
args: []*Device{
{
Index: 0,
nvlibDevice: nvlibDevice{
UUID: "gpu0",
},
Links: map[int][]P2PLink{
1: {{Type: SingleNVLINKLink}},
},
},
{
Index: 1,
nvlibDevice: nvlibDevice{
UUID: "gpu1",
},
Links: map[int][]P2PLink{
0: {{Type: SingleNVLINKLink}, {Type: SingleNVLINKLink}}, // Extra links!
},
},
},
want: ListDeviceScore{
{
UUID: "gpu0",
Score: map[string]int{"gpu1": 100}, // Score for 1 SingleNVLINKLink
},
{
UUID: "gpu1",
Score: map[string]int{"gpu0": 100}, // Evaluates minimum common subset (1 link)
},
},
},
{
name: "one side zero links gracefully handles",
args: []*Device{
{
Index: 0,
nvlibDevice: nvlibDevice{
UUID: "gpu0",
},
Links: map[int][]P2PLink{
1: {}, // Empty slice
},
},
{
Index: 1,
nvlibDevice: nvlibDevice{
UUID: "gpu1",
},
Links: map[int][]P2PLink{
0: {{Type: SingleNVLINKLink}},
},
},
},
want: ListDeviceScore{
{
UUID: "gpu0",
Score: map[string]int{"gpu1": 0},
},
{
UUID: "gpu1",
Score: map[string]int{"gpu0": 0},
},
},
},
{
name: "nil slices gracefully handles missing index",
args: []*Device{
{
Index: 0,
nvlibDevice: nvlibDevice{
UUID: "gpu0",
},
Links: map[int][]P2PLink{
1: nil, // Nil slice
},
},
{
Index: 1,
nvlibDevice: nvlibDevice{
UUID: "gpu1",
},
Links: map[int][]P2PLink{
// gpu0 index 0 is missing entirely from the map
},
},
},
want: ListDeviceScore{
{
UUID: "gpu0",
Score: map[string]int{"gpu1": 0},
},
{
UUID: "gpu1",
Score: map[string]int{"gpu0": 0},
},
},
},
}
for _, test := range tests {
t.Run(test.name, func(t *testing.T) {
Expand Down
Loading