Skip to content
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

GPU sharing on cuda compute capability >=7.5 #231

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

GPU sharing on cuda compute capability >=7.5 #231

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
23 changes: 21 additions & 2 deletions cmd/nvidia-dra-plugin/device_state.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,6 +20,7 @@ import (
"context"
"fmt"
"slices"
"strings"
"sync"

resourceapi "k8s.io/api/resource/v1beta1"
Expand All @@ -29,6 +30,8 @@ import (
"k8s.io/kubernetes/pkg/kubelet/checkpointmanager"
cdiapi "tags.cncf.io/container-device-interface/pkg/cdi"

"golang.org/x/mod/semver"

configapi "github.com/NVIDIA/k8s-dra-driver/api/nvidia.com/resource/gpu/v1alpha1"
)

Expand Down Expand Up @@ -390,6 +393,21 @@ func (s *DeviceState) applySharingConfig(ctx context.Context, config configapi.S
allocatableDevices[r.Device] = s.allocatable[r.Device]
}

// allow devices only with cuda compute compatility >= 7.5 as time slicing and MPS does not work with old arch
shareableAllocatableDevices := make(AllocatableDevices)
for device, deviceType := range allocatableDevices {
if deviceType.Gpu != nil {
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Does this mean that we don't timeslice MIG devices?

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

In general, does it make sense to factor these checks into a function where we can better test the various combinations of options?

cudaCCv := "v" + strings.TrimPrefix(deviceType.Gpu.cudaComputeCapability, "v")
gpuUUID := deviceType.Gpu.UUID
if semver.Compare(semver.Canonical(cudaCCv), semver.Canonical("v7.5")) >= 0 {
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@guptaNswati where does the v7.5 threshold come from? In #58 we check for >= v7.0 and for MPS specifically, v3.5 is mentioned.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I picked it from our device-plugin code checking if its Volta https://github.com/NVIDIA/k8s-device-plugin/blob/main/cmd/mps-control-daemon/mps/device.go#L51

klog.Infof("GPU sharing is available on this device UUID=%v with CudaComputeCapability=%v", gpuUUID, cudaCCv)
shareableAllocatableDevices[device] = deviceType
} else {
return nil, fmt.Errorf("GPU sharing is not available on this device UUID=%v", gpuUUID)
}
}
}

// Declare a device group state object to populate.
var configState DeviceConfigState

Expand All @@ -400,7 +418,7 @@ func (s *DeviceState) applySharingConfig(ctx context.Context, config configapi.S
return nil, fmt.Errorf("error getting timeslice config for requests '%v' in claim '%v': %w", requests, claim.UID, err)
}
if tsc != nil {
err = s.tsManager.SetTimeSlice(allocatableDevices, tsc)
err = s.tsManager.SetTimeSlice(shareableAllocatableDevices, tsc)
if err != nil {
return nil, fmt.Errorf("error setting timeslice config for requests '%v' in claim '%v': %w", requests, claim.UID, err)
}
Expand All @@ -413,7 +431,8 @@ func (s *DeviceState) applySharingConfig(ctx context.Context, config configapi.S
if err != nil {
return nil, fmt.Errorf("error getting MPS configuration: %w", err)
}
mpsControlDaemon := s.mpsManager.NewMpsControlDaemon(string(claim.UID), allocatableDevices)

mpsControlDaemon := s.mpsManager.NewMpsControlDaemon(string(claim.UID), shareableAllocatableDevices)
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Should we distinguish between timeslicing-sharable and MPS-sharable devices?

if err := mpsControlDaemon.Start(ctx, mpsc); err != nil {
return nil, fmt.Errorf("error starting MPS control daemon: %w", err)
}
Expand Down
Loading