github/kubernetes
4
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

WIP v0 NVIDIA GPU support

Browse Source 
Implements part of #24071

I am not familiar with the scheduler enough to know what to do with the scores. Punting for now.

Missing items from the implementation plan: limitranger, rkt support, kubectl
support and user docs
This commit is contained in:
Rudi Chiarito
2016-04-26 20:54:19 -04:00
parent 57359e4f2f
commit 362c763fca
21 changed files with 858 additions and 666 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
plugin/pkg/scheduler/algorithm/predicates/error.go
View File

@@ -19,9 +19,10 @@ package predicates
import "fmt"
const (
podCountResourceName string = "PodCount"
cpuResourceName string = "CPU"
memoryResoureceName string = "Memory"
podCountResourceName string = "PodCount"
cpuResourceName string = "CPU"
memoryResoureceName string = "Memory"
nvidiaGpuResourceName string = "NvidiaGpu"
)
var (

24
plugin/pkg/scheduler/algorithm/predicates/predicates.go
View File

@@ -346,8 +346,9 @@ func (c *VolumeZoneChecker) predicate(pod *api.Pod, nodeInfo *schedulercache.Nod
}
type resourceRequest struct {
milliCPU int64
memory int64
milliCPU int64
memory int64
nvidiaGPU int64
}
func getResourceRequest(pod *api.Pod) resourceRequest {
@@ -356,19 +357,23 @@ func getResourceRequest(pod *api.Pod) resourceRequest {
requests := container.Resources.Requests
result.memory += requests.Memory().Value()
result.milliCPU += requests.Cpu().MilliValue()
result.nvidiaGPU += requests.NvidiaGPU().Value()
}
return result
}
func CheckPodsExceedingFreeResources(pods []*api.Pod, allocatable api.ResourceList) (fitting []*api.Pod, notFittingCPU, notFittingMemory []*api.Pod) {
func CheckPodsExceedingFreeResources(pods []*api.Pod, allocatable api.ResourceList) (fitting []*api.Pod, notFittingCPU, notFittingMemory, notFittingNvidiaGPU []*api.Pod) {
totalMilliCPU := allocatable.Cpu().MilliValue()
totalMemory := allocatable.Memory().Value()
totalNvidiaGPU := allocatable.NvidiaGPU().Value()
milliCPURequested := int64(0)
memoryRequested := int64(0)
nvidiaGPURequested := int64(0)
for _, pod := range pods {
podRequest := getResourceRequest(pod)
fitsCPU := (totalMilliCPU - milliCPURequested) >= podRequest.milliCPU
fitsMemory := (totalMemory - memoryRequested) >= podRequest.memory
fitsNVidiaGPU := (totalNvidiaGPU - nvidiaGPURequested) >= podRequest.nvidiaGPU
if !fitsCPU {
// the pod doesn't fit due to CPU request
notFittingCPU = append(notFittingCPU, pod)
@@ -379,9 +384,15 @@ func CheckPodsExceedingFreeResources(pods []*api.Pod, allocatable api.ResourceLi
notFittingMemory = append(notFittingMemory, pod)
continue
}
if !fitsNVidiaGPU {
// the pod doesn't fit due to NvidiaGPU request
notFittingNvidiaGPU = append(notFittingNvidiaGPU, pod)
continue
}
// the pod fits
milliCPURequested += podRequest.milliCPU
memoryRequested += podRequest.memory
nvidiaGPURequested += podRequest.nvidiaGPU
fitting = append(fitting, pod)
}
return
@@ -403,12 +414,13 @@ func PodFitsResources(pod *api.Pod, nodeInfo *schedulercache.NodeInfo) (bool, er
newInsufficientResourceError(podCountResourceName, 1, int64(len(nodeInfo.Pods())), allowedPodNumber)
}
podRequest := getResourceRequest(pod)
if podRequest.milliCPU == 0 && podRequest.memory == 0 {
if podRequest.milliCPU == 0 && podRequest.memory == 0 && podRequest.nvidiaGPU == 0 {
return true, nil
}
totalMilliCPU := allocatable.Cpu().MilliValue()
totalMemory := allocatable.Memory().Value()
totalNvidiaGPU := allocatable.NvidiaGPU().Value()
if totalMilliCPU < podRequest.milliCPU+nodeInfo.RequestedResource().MilliCPU {
return false,
@@ -418,6 +430,10 @@ func PodFitsResources(pod *api.Pod, nodeInfo *schedulercache.NodeInfo) (bool, er
return false,
newInsufficientResourceError(memoryResoureceName, podRequest.memory, nodeInfo.RequestedResource().Memory, totalMemory)
}
if totalNvidiaGPU < podRequest.nvidiaGPU+nodeInfo.RequestedResource().NvidiaGPU {
return false,
newInsufficientResourceError(nvidiaGpuResourceName, podRequest.nvidiaGPU, nodeInfo.RequestedResource().NvidiaGPU, totalNvidiaGPU)
}
glog.V(10).Infof("Schedule Pod %+v on Node %+v is allowed, Node is running only %v out of %v Pods.",
podName(pod), node.Name, len(nodeInfo.Pods()), allowedPodNumber)
return true, nil

62
plugin/pkg/scheduler/algorithm/predicates/predicates_test.go
View File

@@ -71,21 +71,23 @@ func (pvs FakePersistentVolumeInfo) GetPersistentVolumeInfo(pvID string) (*api.P
return nil, fmt.Errorf("Unable to find persistent volume: %s", pvID)
}
func makeResources(milliCPU int64, memory int64, pods int64) api.NodeResources {
func makeResources(milliCPU int64, memory int64, nvidiaGPUs int64, pods int64) api.NodeResources {
return api.NodeResources{
Capacity: api.ResourceList{
api.ResourceCPU: *resource.NewMilliQuantity(milliCPU, resource.DecimalSI),
api.ResourceMemory: *resource.NewQuantity(memory, resource.BinarySI),
api.ResourcePods: *resource.NewQuantity(pods, resource.DecimalSI),
api.ResourceCPU: *resource.NewMilliQuantity(milliCPU, resource.DecimalSI),
api.ResourceMemory: *resource.NewQuantity(memory, resource.BinarySI),
api.ResourcePods: *resource.NewQuantity(pods, resource.DecimalSI),
api.ResourceNvidiaGPU: *resource.NewQuantity(nvidiaGPUs, resource.DecimalSI),
},
}
}
func makeAllocatableResources(milliCPU int64, memory int64, pods int64) api.ResourceList {
func makeAllocatableResources(milliCPU int64, memory int64, nvidiaGPUs int64, pods int64) api.ResourceList {
return api.ResourceList{
api.ResourceCPU: *resource.NewMilliQuantity(milliCPU, resource.DecimalSI),
api.ResourceMemory: *resource.NewQuantity(memory, resource.BinarySI),
api.ResourcePods: *resource.NewQuantity(pods, resource.DecimalSI),
api.ResourceCPU: *resource.NewMilliQuantity(milliCPU, resource.DecimalSI),
api.ResourceMemory: *resource.NewQuantity(memory, resource.BinarySI),
api.ResourcePods: *resource.NewQuantity(pods, resource.DecimalSI),
api.ResourceNvidiaGPU: *resource.NewQuantity(nvidiaGPUs, resource.DecimalSI),
}
}
@@ -95,8 +97,9 @@ func newResourcePod(usage ...resourceRequest) *api.Pod {
containers = append(containers, api.Container{
Resources: api.ResourceRequirements{
Requests: api.ResourceList{
api.ResourceCPU: *resource.NewMilliQuantity(req.milliCPU, resource.DecimalSI),
api.ResourceMemory: *resource.NewQuantity(req.memory, resource.BinarySI),
api.ResourceCPU: *resource.NewMilliQuantity(req.milliCPU, resource.DecimalSI),
api.ResourceMemory: *resource.NewQuantity(req.memory, resource.BinarySI),
api.ResourceNvidiaGPU: *resource.NewQuantity(req.nvidiaGPU, resource.DecimalSI),
},
},
})
@@ -159,7 +162,7 @@ func TestPodFitsResources(t *testing.T) {
}
for _, test := range enoughPodsTests {
node := api.Node{Status: api.NodeStatus{Capacity: makeResources(10, 20, 32).Capacity, Allocatable: makeAllocatableResources(10, 20, 32)}}
node := api.Node{Status: api.NodeStatus{Capacity: makeResources(10, 20, 0, 32).Capacity, Allocatable: makeAllocatableResources(10, 20, 0, 32)}}
test.nodeInfo.SetNode(&node)
fits, err := PodFitsResources(test.pod, test.nodeInfo)
@@ -204,7 +207,7 @@ func TestPodFitsResources(t *testing.T) {
},
}
for _, test := range notEnoughPodsTests {
node := api.Node{Status: api.NodeStatus{Capacity: api.ResourceList{}, Allocatable: makeAllocatableResources(10, 20, 1)}}
node := api.Node{Status: api.NodeStatus{Capacity: api.ResourceList{}, Allocatable: makeAllocatableResources(10, 20, 0, 1)}}
test.nodeInfo.SetNode(&node)
fits, err := PodFitsResources(test.pod, test.nodeInfo)
@@ -1529,7 +1532,7 @@ func TestRunGeneralPredicates(t *testing.T) {
newResourcePod(resourceRequest{milliCPU: 9, memory: 19})),
node: &api.Node{
ObjectMeta: api.ObjectMeta{Name: "machine1"},
Status: api.NodeStatus{Capacity: makeResources(10, 20, 32).Capacity, Allocatable: makeAllocatableResources(10, 20, 32)},
Status: api.NodeStatus{Capacity: makeResources(10, 20, 0, 32).Capacity, Allocatable: makeAllocatableResources(10, 20, 0, 32)},
},
fits: true,
wErr: nil,
@@ -1541,12 +1544,39 @@ func TestRunGeneralPredicates(t *testing.T) {
newResourcePod(resourceRequest{milliCPU: 5, memory: 19})),
node: &api.Node{
ObjectMeta: api.ObjectMeta{Name: "machine1"},
Status: api.NodeStatus{Capacity: makeResources(10, 20, 32).Capacity, Allocatable: makeAllocatableResources(10, 20, 32)},
Status: api.NodeStatus{Capacity: makeResources(10, 20, 0, 32).Capacity, Allocatable: makeAllocatableResources(10, 20, 0, 32)},
},
fits: false,
wErr: newInsufficientResourceError("CPU", 8, 5, 10),
test: "not enough cpu resource",
},
{
pod: &api.Pod{},
nodeInfo: schedulercache.NewNodeInfo(
newResourcePod(resourceRequest{milliCPU: 9, memory: 19})),
node: &api.Node{Status: api.NodeStatus{Capacity: makeResources(10, 20, 1, 32).Capacity, Allocatable: makeAllocatableResources(10, 20, 1, 32)}},
fits: true,
wErr: nil,
test: "no resources/port/host requested always fits on GPU machine",
},
{
pod: newResourcePod(resourceRequest{milliCPU: 3, memory: 1, nvidiaGPU: 1}),
nodeInfo: schedulercache.NewNodeInfo(
newResourcePod(resourceRequest{milliCPU: 5, memory: 10, nvidiaGPU: 1})),
node: &api.Node{Status: api.NodeStatus{Capacity: makeResources(10, 20, 1, 32).Capacity, Allocatable: makeAllocatableResources(10, 20, 1, 32)}},
fits: false,
wErr: newInsufficientResourceError("NvidiaGpu", 1, 1, 1),
test: "not enough GPU resource",
},
{
pod: newResourcePod(resourceRequest{milliCPU: 3, memory: 1, nvidiaGPU: 1}),
nodeInfo: schedulercache.NewNodeInfo(
newResourcePod(resourceRequest{milliCPU: 5, memory: 10, nvidiaGPU: 0})),
node: &api.Node{Status: api.NodeStatus{Capacity: makeResources(10, 20, 1, 32).Capacity, Allocatable: makeAllocatableResources(10, 20, 1, 32)}},
fits: true,
wErr: nil,
test: "enough GPU resource",
},
{
pod: &api.Pod{
Spec: api.PodSpec{
@@ -1556,7 +1586,7 @@ func TestRunGeneralPredicates(t *testing.T) {
nodeInfo: schedulercache.NewNodeInfo(),
node: &api.Node{
ObjectMeta: api.ObjectMeta{Name: "machine1"},
Status: api.NodeStatus{Capacity: makeResources(10, 20, 32).Capacity, Allocatable: makeAllocatableResources(10, 20, 32)},
Status: api.NodeStatus{Capacity: makeResources(10, 20, 0, 32).Capacity, Allocatable: makeAllocatableResources(10, 20, 0, 32)},
},
fits: false,
wErr: ErrPodNotMatchHostName,
@@ -1567,7 +1597,7 @@ func TestRunGeneralPredicates(t *testing.T) {
nodeInfo: schedulercache.NewNodeInfo(newPodWithPort(123)),
node: &api.Node{
ObjectMeta: api.ObjectMeta{Name: "machine1"},
Status: api.NodeStatus{Capacity: makeResources(10, 20, 32).Capacity, Allocatable: makeAllocatableResources(10, 20, 32)},
Status: api.NodeStatus{Capacity: makeResources(10, 20, 0, 32).Capacity, Allocatable: makeAllocatableResources(10, 20, 0, 32)},
},
fits: false,
wErr: ErrPodNotFitsHostPorts,