New scheduler. BinPacking with capping each node to different values (piston capping).
This commit is contained in:
Pradyumna Kaushik
Renan DelValle
parent
a6dcda04c8
commit
d9357f59e7
1 changed files with 409 additions and 0 deletions
409
schedulers/pistoncapper.go
Normal file
409
schedulers/pistoncapper.go
Normal file
|
|
@ -0,0 +1,409 @@
|
|||
package schedulers
|
||||
|
||||
import (
|
||||
"bitbucket.org/sunybingcloud/electron/constants"
|
||||
"bitbucket.org/sunybingcloud/electron/def"
|
||||
"bitbucket.org/sunybingcloud/electron/rapl"
|
||||
"fmt"
|
||||
"errors"
|
||||
"github.com/golang/protobuf/proto"
|
||||
mesos "github.com/mesos/mesos-go/mesosproto"
|
||||
"github.com/mesos/mesos-go/mesosutil"
|
||||
sched "github.com/mesos/mesos-go/scheduler"
|
||||
"log"
|
||||
"math"
|
||||
"sync"
|
||||
"strings"
|
||||
"time"
|
||||
)
|
||||
|
||||
/*
|
||||
Piston Capper implements the Scheduler interface
|
||||
|
||||
This is basically extending the BinPacking algorithm to also cap each node at a different values,
|
||||
corresponding to the load on that node.
|
||||
*/
|
||||
type PistonCapper struct {
|
||||
tasksCreated int
|
||||
tasksRunning int
|
||||
tasks []def.Task
|
||||
metrics map[string]def.Metric
|
||||
running map[string]map[string]bool
|
||||
taskMonitor map[string][]def.Task
|
||||
clusterLoad map[string]float64
|
||||
totalPower map[string]float64
|
||||
ignoreWatts bool
|
||||
ticker *time.Ticker
|
||||
isCapping bool
|
||||
|
||||
// First set of PCP values are garbage values, signal to logger to start recording when we're
|
||||
// about to schedule the new task.
|
||||
RecordPCP bool
|
||||
|
||||
// This channel is closed when the program receives an interrupt,
|
||||
// signalling that the program should shut down.
|
||||
Shutdown chan struct{}
|
||||
|
||||
// This channel is closed after shutdown is closed, and only when all
|
||||
// outstanding tasks have been cleaned up.
|
||||
Done chan struct{}
|
||||
|
||||
// Controls when to shutdown pcp logging.
|
||||
PCPLog chan struct{}
|
||||
}
|
||||
|
||||
// New electron scheduler.
|
||||
func NewPistonCapper(tasks []def.Task, ignoreWatts bool) *PistonCapper {
|
||||
s := &PistonCapper{
|
||||
tasks: tasks,
|
||||
ignoreWatts: ignoreWatts,
|
||||
Shutdown: make(chan struct{}),
|
||||
Done: make(chan struct{}),
|
||||
PCPLog: make(chan struct{}),
|
||||
running: make(map[string]map[string]bool),
|
||||
taskMonitor: make(map[string][]def.Task),
|
||||
clusterLoad: make(map[string]float64),
|
||||
totalPower: make(map[string]float64),
|
||||
RecordPCP: false,
|
||||
ticker: time.NewTicker(10 * time.Second),
|
||||
isCapping: false,
|
||||
}
|
||||
return s
|
||||
}
|
||||
|
||||
// mutex
|
||||
var mutex sync.Mutex
|
||||
|
||||
func (s *PistonCapper) newTask(offer *mesos.Offer, task def.Task) *mesos.TaskInfo {
|
||||
taskName := fmt.Sprintf("%s-%d", task.Name, *task.Instances)
|
||||
s.tasksCreated++
|
||||
|
||||
if !s.RecordPCP {
|
||||
// Turn on logging
|
||||
s.RecordPCP = true
|
||||
time.Sleep(1 * time.Second) // Make sure we're recording by the time the first task starts
|
||||
}
|
||||
|
||||
// If this is our first time running into this Agent
|
||||
if _, ok := s.running[offer.GetSlaveId().GoString()]; !ok {
|
||||
s.running[offer.GetSlaveId().GoString()] = make(map[string]bool)
|
||||
}
|
||||
|
||||
// Setting the task ID to the task. This is done so that we can consider each task to be different,
|
||||
// even though they have the same parameters.
|
||||
task.SetTaskID(*proto.String("electron-" + taskName))
|
||||
// Add task to list of tasks running on node
|
||||
s.running[offer.GetSlaveId().GoString()][taskName] = true
|
||||
// Adding the task to the taskMonitor
|
||||
if len(s.taskMonitor[*offer.Hostname]) == 0 {
|
||||
s.taskMonitor[*offer.Hostname] = []def.Task{task}
|
||||
} else {
|
||||
s.taskMonitor[*offer.Hostname] = append(s.taskMonitor[*offer.Hostname], task)
|
||||
}
|
||||
|
||||
resources := []*mesos.Resource{
|
||||
mesosutil.NewScalarResource("cpus", task.CPU),
|
||||
mesosutil.NewScalarResource("mem", task.RAM),
|
||||
}
|
||||
|
||||
if !s.ignoreWatts {
|
||||
resources = append(resources, mesosutil.NewScalarResource("watts", task.Watts))
|
||||
}
|
||||
|
||||
return &mesos.TaskInfo{
|
||||
Name: proto.String(taskName),
|
||||
TaskId: &mesos.TaskID{
|
||||
Value: proto.String("electron-" + taskName),
|
||||
},
|
||||
SlaveId: offer.SlaveId,
|
||||
Resources: resources,
|
||||
Command: &mesos.CommandInfo{
|
||||
Value: proto.String(task.CMD),
|
||||
},
|
||||
Container: &mesos.ContainerInfo{
|
||||
Type: mesos.ContainerInfo_DOCKER.Enum(),
|
||||
Docker: &mesos.ContainerInfo_DockerInfo{
|
||||
Image: proto.String(task.Image),
|
||||
Network: mesos.ContainerInfo_DockerInfo_BRIDGE.Enum(), // Run everything isolated
|
||||
},
|
||||
},
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
|
||||
func (s *PistonCapper) Registered(
|
||||
_ sched.SchedulerDriver,
|
||||
frameworkID *mesos.FrameworkID,
|
||||
masterInfo *mesos.MasterInfo) {
|
||||
log.Printf("Framework %s registered with master %s", frameworkID, masterInfo)
|
||||
}
|
||||
|
||||
func (s *PistonCapper) Reregistered(_ sched.SchedulerDriver, masterInfo *mesos.MasterInfo) {
|
||||
log.Printf("Framework re-registered with master %s", masterInfo)
|
||||
}
|
||||
|
||||
func (s *PistonCapper) Disconnected(sched.SchedulerDriver) {
|
||||
log.Println("Framework disconnected with master")
|
||||
}
|
||||
|
||||
// go routine to cap the each node in the cluster at regular intervals of time.
|
||||
var capValues = make(map[string]float64)
|
||||
// Storing the previous cap value for each host so as to not repeatedly cap the nodes to the same value. (reduces overhead)
|
||||
var previousRoundedCapValues = make(map[string]int)
|
||||
func (s *PistonCapper) startCapping() {
|
||||
go func() {
|
||||
for {
|
||||
select {
|
||||
case <-s.ticker.C:
|
||||
// Need to cap each node
|
||||
mutex.Lock()
|
||||
for host, capValue := range capValues {
|
||||
roundedCapValue := int(math.Floor(capValue + 0.5))
|
||||
// has the cap value changed
|
||||
if prevRoundedCap, ok := previousRoundedCapValues[host]; ok {
|
||||
if prevRoundedCap != roundedCapValue {
|
||||
if err := rapl.Cap(host, "rapl", roundedCapValue); err != nil {
|
||||
log.Println(err)
|
||||
}
|
||||
log.Printf("Capped [%s] at %d", host, int(math.Floor(capValue + 0.5)))
|
||||
previousRoundedCapValues[host] = roundedCapValue
|
||||
}
|
||||
} else {
|
||||
if err := rapl.Cap(host, "rapl", roundedCapValue); err != nil {
|
||||
log.Println(err)
|
||||
}
|
||||
log.Printf("Capped [%s] at %d", host, int(math.Floor(capValue + 0.5)))
|
||||
previousRoundedCapValues[host] = roundedCapValue
|
||||
}
|
||||
}
|
||||
mutex.Unlock()
|
||||
}
|
||||
}
|
||||
}()
|
||||
}
|
||||
|
||||
// Stop the capping
|
||||
func (s *PistonCapper) stopCapping() {
|
||||
if s.isCapping {
|
||||
log.Println("Stopping the capping.")
|
||||
s.ticker.Stop()
|
||||
mutex.Lock()
|
||||
s.isCapping = false
|
||||
mutex.Unlock()
|
||||
}
|
||||
}
|
||||
|
||||
// Check whether we are overloading the host (from watts perspective)
|
||||
func wattsOverload(task def.Task, offerWatts float64, totalPower float64) bool {
|
||||
if offerWatts >= (totalPower + (task.Watts * constants.CapMargin)) {
|
||||
return false
|
||||
} else {
|
||||
return true
|
||||
}
|
||||
}
|
||||
|
||||
func (s *PistonCapper) ResourceOffers(driver sched.SchedulerDriver, offers []*mesos.Offer) {
|
||||
log.Printf("Received %d resource offers", len(offers))
|
||||
|
||||
// retrieving the total power for each host in the offers
|
||||
for _, offer := range offers {
|
||||
if _, ok := s.totalPower[*offer.Hostname]; !ok {
|
||||
_, _, offer_watts := OfferAgg(offer)
|
||||
s.totalPower[*offer.Hostname] = offer_watts
|
||||
}
|
||||
}
|
||||
|
||||
// Displaying the totalPower
|
||||
for host, tpower := range s.totalPower {
|
||||
log.Printf("TotalPower[%s] = %f", host, tpower)
|
||||
}
|
||||
|
||||
/*
|
||||
Piston capping strategy
|
||||
|
||||
Perform bin-packing of tasks on nodes in the cluster, making sure that no task is given less hard-limit resources than requested.
|
||||
For each set of tasks that are scheduled, compute the new cap values for each host in the cluster.
|
||||
At regular intervals of time, cap each node in the cluster.
|
||||
*/
|
||||
log.Printf("Number of tasks yet to be scheduled: %d", len(s.tasks))
|
||||
|
||||
|
||||
for _, offer := range offers {
|
||||
select {
|
||||
case <-s.Shutdown:
|
||||
log.Println("Done scheduling tasks: declining offer on [", offer.GetHostname(), "]")
|
||||
driver.DeclineOffer(offer.Id, longFilter)
|
||||
|
||||
log.Println("Number of tasks still running: ", s.tasksRunning)
|
||||
continue
|
||||
default:
|
||||
}
|
||||
|
||||
fitTasks := []*mesos.TaskInfo{}
|
||||
offerCPU, offerRAM, offerWatts := OfferAgg(offer)
|
||||
taken := false
|
||||
totalWatts := 0.0
|
||||
totalCPU := 0.0
|
||||
totalRAM := 0.0
|
||||
// Store the partialLoad for host corresponding to this offer.
|
||||
// Once we can't fit any more tasks, we update capValue for this host with partialLoad and then launch the fit tasks.
|
||||
partialLoad := 0.0
|
||||
for i, task := range s.tasks {
|
||||
// Check host if it exists
|
||||
if task.Host != "" {
|
||||
// Don't take offer if it doens't match our task's host requirement.
|
||||
if !strings.HasPrefix(*offer.Hostname, task.Host) {
|
||||
continue
|
||||
}
|
||||
}
|
||||
|
||||
for *task.Instances > 0 {
|
||||
// Does the task fit
|
||||
if (s.ignoreWatts || !wattsOverload(task, offerWatts, totalWatts)) &&
|
||||
(offerCPU >= (totalCPU + task.CPU)) &&
|
||||
(offerRAM >= (totalRAM + task.RAM)) {
|
||||
|
||||
// Start piston capping if haven't started yet
|
||||
if !s.isCapping {
|
||||
s.isCapping = true
|
||||
s.startCapping()
|
||||
}
|
||||
|
||||
taken = true
|
||||
totalWatts += (task.Watts * constants.CapMargin)
|
||||
totalCPU += task.CPU
|
||||
totalRAM += task.RAM
|
||||
log.Println("Co-Located with: ")
|
||||
coLocated(s.running[offer.GetSlaveId().GoString()])
|
||||
fitTasks = append(fitTasks, s.newTask(offer, task))
|
||||
|
||||
log.Println("Inst: ", *task.Instances)
|
||||
*task.Instances--
|
||||
// updating the cap value for offer.Hostname
|
||||
partialLoad += ((task.Watts * constants.CapMargin) / s.totalPower[*offer.Hostname]) * 100
|
||||
|
||||
if *task.Instances <= 0 {
|
||||
// All instances of task have been scheduled. Remove it
|
||||
s.tasks = append(s.tasks[:i], s.tasks[i+1:]...)
|
||||
if len(s.tasks) <= 0 {
|
||||
log.Println("Done scheduling all tasks")
|
||||
close(s.Shutdown)
|
||||
}
|
||||
}
|
||||
} else {
|
||||
break // Continue on to next task
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if taken {
|
||||
// Updating the cap value for offer.Hostname
|
||||
mutex.Lock()
|
||||
capValues[*offer.Hostname] += partialLoad
|
||||
mutex.Unlock()
|
||||
log.Printf("Starting on [%s]\n", offer.GetHostname())
|
||||
driver.LaunchTasks([]*mesos.OfferID{offer.Id}, fitTasks, defaultFilter)
|
||||
} else {
|
||||
// If there was no match for task
|
||||
log.Println("There is not enough resources to launch task: ")
|
||||
cpus, mem, watts := OfferAgg(offer)
|
||||
|
||||
log.Printf("<CPU: %f, RAM: %f, Watts: %f>\n", cpus, mem, watts)
|
||||
driver.DeclineOffer(offer.Id, defaultFilter)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Remove finished task from the taskMonitor
|
||||
func (s *PistonCapper) deleteFromTaskMonitor(finishedTaskID string) (def.Task, string, error) {
|
||||
hostOfFinishedTask := ""
|
||||
indexOfFinishedTask := -1
|
||||
found := false
|
||||
var finishedTask def.Task
|
||||
|
||||
for host, tasks := range s.taskMonitor {
|
||||
for i, task := range tasks {
|
||||
if task.TaskID == finishedTaskID {
|
||||
hostOfFinishedTask = host
|
||||
indexOfFinishedTask = i
|
||||
found = true
|
||||
}
|
||||
}
|
||||
if found {
|
||||
break
|
||||
}
|
||||
}
|
||||
|
||||
if hostOfFinishedTask != "" && indexOfFinishedTask != -1 {
|
||||
finishedTask = s.taskMonitor[hostOfFinishedTask][indexOfFinishedTask]
|
||||
log.Printf("Removing task with TaskID [%s] from the list of running tasks\n",
|
||||
s.taskMonitor[hostOfFinishedTask][indexOfFinishedTask].TaskID)
|
||||
s.taskMonitor[hostOfFinishedTask] = append(s.taskMonitor[hostOfFinishedTask][:indexOfFinishedTask],
|
||||
s.taskMonitor[hostOfFinishedTask][indexOfFinishedTask+1:]...)
|
||||
} else {
|
||||
return finishedTask, hostOfFinishedTask, errors.New("Finished Task not present in TaskMonitor")
|
||||
}
|
||||
return finishedTask, hostOfFinishedTask, nil
|
||||
}
|
||||
|
||||
func (s *PistonCapper) StatusUpdate(driver sched.SchedulerDriver, status *mesos.TaskStatus) {
|
||||
log.Printf("Received task status [%s] for task [%s]\n", NameFor(status.State), *status.TaskId.Value)
|
||||
|
||||
if *status.State == mesos.TaskState_TASK_RUNNING {
|
||||
mutex.Lock()
|
||||
s.tasksRunning++
|
||||
mutex.Unlock()
|
||||
} else if IsTerminal(status.State) {
|
||||
delete(s.running[status.GetSlaveId().GoString()], *status.TaskId.Value)
|
||||
// Deleting the task from the taskMonitor
|
||||
finishedTask, hostOfFinishedTask, err := s.deleteFromTaskMonitor(*status.TaskId.Value)
|
||||
if err != nil {
|
||||
log.Println(err)
|
||||
}
|
||||
|
||||
// Need to update the cap values for host of the finishedTask
|
||||
mutex.Lock()
|
||||
capValues[hostOfFinishedTask] -= ((finishedTask.Watts * constants.CapMargin) / s.totalPower[hostOfFinishedTask]) * 100
|
||||
// Checking to see if the cap value has become 0, in which case we uncap the host.
|
||||
if int(math.Floor(capValues[hostOfFinishedTask] + 0.5)) == 0 {
|
||||
capValues[hostOfFinishedTask] = 100
|
||||
}
|
||||
s.tasksRunning--
|
||||
mutex.Unlock()
|
||||
|
||||
if s.tasksRunning == 0 {
|
||||
select {
|
||||
case <-s.Shutdown:
|
||||
s.stopCapping()
|
||||
close(s.Done)
|
||||
default:
|
||||
}
|
||||
}
|
||||
}
|
||||
log.Printf("DONE: Task status [%s] for task [%s]", NameFor(status.State), *status.TaskId.Value)
|
||||
}
|
||||
|
||||
func (s *PistonCapper) FrameworkMessage(
|
||||
driver sched.SchedulerDriver,
|
||||
executorID *mesos.ExecutorID,
|
||||
slaveID *mesos.SlaveID,
|
||||
message string) {
|
||||
|
||||
log.Println("Getting a framework message: ", message)
|
||||
log.Printf("Received a framework message from some unknown source: %s", *executorID.Value)
|
||||
}
|
||||
|
||||
func (s *PistonCapper) OfferRescinded(_ sched.SchedulerDriver, offerID *mesos.OfferID) {
|
||||
log.Printf("Offer %s rescinded", offerID)
|
||||
}
|
||||
func (s *PistonCapper) SlaveLost(_ sched.SchedulerDriver, slaveID *mesos.SlaveID) {
|
||||
log.Printf("Slave %s lost", slaveID)
|
||||
}
|
||||
func (s *PistonCapper) ExecutorLost(_ sched.SchedulerDriver, executorID *mesos.ExecutorID, slaveID *mesos.SlaveID, status int) {
|
||||
log.Printf("Executor %s on slave %s was lost", executorID, slaveID)
|
||||
}
|
||||
|
||||
func (s *PistonCapper) Error(_ sched.SchedulerDriver, err string) {
|
||||
log.Printf("Receiving an error: %s", err)
|
||||
}
|
||||
Reference in a new issue