From 6bf9795d5e0dfc705299381dc902b22d03ded063 Mon Sep 17 00:00:00 2001 From: Tao Wang Date: Tue, 25 Jan 2022 19:23:43 -0500 Subject: [PATCH] kubernetes: make isolcpus allocation SMT-aware Enhance isolcpus support in Kubernetes to allocate isolated SMT siblings to the same container when SMT/HT is enabled on the host. As it stands, the device manager code in Kubernetes is not SMT-aware (since normally it doesn't deal with CPUs). However, StarlingX exposes isolated CPUs as devices and if possible we want to allocate all SMT siblings from a CPU core to the same container in order to minimize cross- container interference due to resource contention within the CPU core. The solution is basically to take the list of isolated CPUs and re-order it so that the SMT siblings are next to each other. That way the existing resource selection code will allocate the siblings together. As an optimization, if it is known that an odd number of isolated CPUs are desired, a singleton SMT sibling will be inserted into the list to avoid breaking up sibling pairs. Signed-off-by: Tao Wang --- pkg/kubelet/cm/devicemanager/manager.go | 153 ++++++++++++++++++++++-- 1 file changed, 146 insertions(+), 7 deletions(-) diff --git a/pkg/kubelet/cm/devicemanager/manager.go b/pkg/kubelet/cm/devicemanager/manager.go index 65b91393..76a0ea6e 100644 --- a/pkg/kubelet/cm/devicemanager/manager.go +++ b/pkg/kubelet/cm/devicemanager/manager.go @@ -19,10 +19,13 @@ package devicemanager import ( "context" "fmt" + "io/ioutil" "net" "os" "path/filepath" "sort" + "strconv" + "strings" "sync" "time" @@ -41,6 +44,7 @@ import ( "k8s.io/kubernetes/pkg/kubelet/checkpointmanager" "k8s.io/kubernetes/pkg/kubelet/checkpointmanager/errors" cputopology "k8s.io/kubernetes/pkg/kubelet/cm/cpumanager/topology" + "k8s.io/kubernetes/pkg/kubelet/cm/cpuset" "k8s.io/kubernetes/pkg/kubelet/cm/devicemanager/checkpoint" "k8s.io/kubernetes/pkg/kubelet/cm/topologymanager" "k8s.io/kubernetes/pkg/kubelet/cm/topologymanager/bitmask" @@ -635,6 +639,80 @@ func (m *ManagerImpl) UpdateAllocatedDevices() { m.allocatedDevices = m.podDevices.devices() } +//Given a list of isolated CPUs in 'devices', and the number of desired CPUs in 'needed', +//return an ordered list of isolated CPUs such that the first 'needed' CPUs in the list +//contain as many hyperthread sibling pairs as possible. +func order_devices_by_sibling(devices sets.String, needed int) ([]string, error) { + var dev_lst []string + var single_lst []string + sibling_lst := make([]string, 0, int(devices.Len())) + _iterated_cpu := make(map[string]string) + + get_sibling := func(cpu string, cpu_lst []string) string { + if cpu_lst[0] == cpu { + return cpu_lst[1] + } else { + return cpu_lst[0] + } + } + for cpu_id := range devices { + // If we've already found cpu_id as a sibling, skip it. + if _, ok := _iterated_cpu[cpu_id]; ok { + continue + } + devPath := fmt.Sprintf("/sys/devices/system/cpu/cpu%s/topology/thread_siblings_list", cpu_id) + dat, err := ioutil.ReadFile(devPath) + if err != nil { + return dev_lst, fmt.Errorf("Can't read cpu[%s] thread_siblings_list", cpu_id) + } + cpustring := strings.TrimSuffix(string(dat), "\n") + cpu_pair_set, err := cpuset.Parse(cpustring) + if err != nil { + return dev_lst, fmt.Errorf("Unable to parse thread_siblings_list[%s] string to cpuset", cpustring) + } + var cpu_pair_lst []string + for _, v := range cpu_pair_set.ToSlice() { + cpu_pair_lst = append(cpu_pair_lst, strconv.Itoa(v)) + } + sibling_cpu_id := get_sibling(cpu_id, cpu_pair_lst) + + if _, ok := devices[sibling_cpu_id]; ok { + sibling_lst = append(sibling_lst, cpu_id, sibling_cpu_id) + _iterated_cpu[sibling_cpu_id] = "" + } else { + single_lst = append(single_lst, cpu_id) + } + _iterated_cpu[cpu_id] = "" + } + + if needed%2 == 0 { + dev_lst = append(sibling_lst, single_lst...) + } else { + if len(single_lst) > 1 { + _tmp_list := append(sibling_lst, single_lst[1:]...) + dev_lst = append(single_lst[0:1], _tmp_list...) + } else { + if len(single_lst) == 0 { + dev_lst = sibling_lst + } else { + dev_lst = append(single_lst, sibling_lst...) + } + + } + } + //klog.Infof("needed=%d ordered_cpu_list=%v", needed, dev_lst) + return dev_lst, nil +} + +func smt_enabled() bool { + dat, _ := ioutil.ReadFile("/sys/devices/system/cpu/smt/active") + state := strings.TrimSuffix(string(dat), "\n") + if state == "0" { + return false + } + return true +} + // Returns list of device Ids we need to allocate with Allocate rpc call. // Returns empty list in case we don't need to issue the Allocate rpc call. func (m *ManagerImpl) devicesToAllocate(podUID, contName, resource string, required int, reusableDevices sets.String) (sets.String, error) { @@ -664,13 +742,29 @@ func (m *ManagerImpl) devicesToAllocate(podUID, contName, resource string, requi } devices = sets.NewString() // Allocates from reusableDevices list first. - for device := range reusableDevices { - devices.Insert(device) - needed-- - if needed == 0 { - return devices, nil + if resource == "windriver.com/isolcpus" && smt_enabled() { + _reusableDevices, err := order_devices_by_sibling(reusableDevices, needed) + if err != nil { + klog.Errorf("error in order_devices_by_sibling: %v", err) + } + // _reusableDevices is type of slice,So we need a unique loop to process here. + for _, device := range _reusableDevices { + devices.Insert(device) + needed-- + if needed == 0 { + return devices, nil + } + } + } else { + for device := range reusableDevices { + devices.Insert(device) + needed-- + if needed == 0 { + return devices, nil + } } } + // Needs to allocate additional devices. if m.allocatedDevices[resource] == nil { m.allocatedDevices[resource] = sets.NewString() @@ -682,13 +776,25 @@ func (m *ManagerImpl) devicesToAllocate(podUID, contName, resource string, requi if available.Len() < needed { return nil, fmt.Errorf("requested number of devices unavailable for %s. Requested: %d, Available: %d", resource, needed, available.Len()) } - // By default, pull devices from the unsorted list of available devices. - allocated := available.UnsortedList()[:needed] + // If topology alignment is desired, update allocated to the set of devices // with the best alignment. + var allocated []string hint := m.topologyAffinityStore.GetAffinity(podUID, contName) if m.deviceHasTopologyAlignment(resource) && hint.NUMANodeAffinity != nil { allocated = m.takeByTopology(resource, available, hint.NUMANodeAffinity, needed) + } else { + if resource == "windriver.com/isolcpus" && smt_enabled() { + var err error + allocated, err = order_devices_by_sibling(available, needed) + allocated = allocated[:needed] + if err != nil { + klog.Errorf("error in order_devices_by_sibling: %v", err) + } + } else { + // By default, pull devices from the unsorted list of available devices. + allocated = available.UnsortedList()[:needed] + } } // Updates m.allocatedDevices with allocated devices to prevent them // from being allocated to other pods/containers, given that we are @@ -764,6 +870,39 @@ func (m *ManagerImpl) takeByTopology(resource string, available sets.String, aff } } + //Add specific logic to process isolcpus resource. + //Try to not sabotage the original logical structure. + //Sort the original three lists by sibling: fromAffinity,notFromAffinity,withoutTopology + if resource == "windriver.com/isolcpus" && smt_enabled() { + var err error + _request_device_map := make(sets.String) + for _, dev := range fromAffinity { + _request_device_map[dev] = sets.Empty{} + } + fromAffinity, err = order_devices_by_sibling(_request_device_map, request) + if err != nil { + klog.Errorf("error in order_devices_by_sibling: %v", err) + } + + _request_device_map = make(sets.String) + for _, dev := range notFromAffinity { + _request_device_map[dev] = sets.Empty{} + } + notFromAffinity, err = order_devices_by_sibling(_request_device_map, request) + if err != nil { + klog.Errorf("error in order_devices_by_sibling: %v", err) + } + + _request_device_map = make(sets.String) + for _, dev := range withoutTopology { + _request_device_map[dev] = sets.Empty{} + } + withoutTopology, err = order_devices_by_sibling(_request_device_map, request) + if err != nil { + klog.Errorf("error in order_devices_by_sibling: %v", err) + } + } + // Concatenate the lists above return the first 'request' devices from it.. return append(append(fromAffinity, notFromAffinity...), withoutTopology...)[:request] } -- 2.22.5