Merged in bpswClassToWattsCapping (pull request #6)

Bpswclasstowattscapping

pcp/proactiveclusterwidecappers.go

@@ -99,7 +99,7 @@ func (capper ClusterwideCapper) CleverRecap(totalPower map[string]float64,
 
 	// determining the Recap value by calling the regular Recap(...)
 	toggle := false
-	RecapValue, err := capper.Recap(totalPower, taskMonitor, finishedTaskId)
+	RecapValue, err := capper.NaiveRecap(totalPower, taskMonitor, finishedTaskId)
 	if err == nil {
 		toggle = true
 	}
@@ -181,7 +181,7 @@ Recapping the entire cluster.
 
 This needs to be called whenever a task finishes execution.
 */
-func (capper ClusterwideCapper) Recap(totalPower map[string]float64,
+func (capper ClusterwideCapper) NaiveRecap(totalPower map[string]float64,
 	taskMonitor map[string][]def.Task, finishedTaskId string) (float64, error) {
 	// Validation
 	if totalPower == nil || taskMonitor == nil {

schedulers/README.md

@@ -7,6 +7,7 @@ To Do:
  * Fix the race condition on 'tasksRunning' in proactiveclusterwidecappingfcfs.go and proactiveclusterwidecappingranked.go
  * Separate the capping strategies from the scheduling algorithms and make it possible to use any capping strategy with any scheduler.
  * Make newTask(...) variadic where the newTaskClass argument can either be given or not. If not give, then pick task.Watts as the watts attribute, else pick task.ClassToWatts[newTaskClass].
+ * Retrofit pcp/proactiveclusterwidecappers.go to include the power capping go routines and to cap only when necessary.
 
 Scheduling Algorithms:
 

schedulers/bpMaxMinPistonCapping.go

@@ -0,0 +1,427 @@
+package schedulers
+
+import (
+	"bitbucket.org/sunybingcloud/electron/constants"
+	"bitbucket.org/sunybingcloud/electron/def"
+	"bitbucket.org/sunybingcloud/electron/rapl"
+	"errors"
+	"fmt"
+	"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"
+	"os"
+	"sort"
+	"strings"
+	"sync"
+	"time"
+)
+
+// Decides if to take an offer or not
+func (s *BPMaxMinPistonCapping) takeOffer(offer *mesos.Offer, task def.Task) bool {
+
+	cpus, mem, watts := OfferAgg(offer)
+
+	//TODO: Insert watts calculation here instead of taking them as a parameter
+
+	if cpus >= task.CPU && mem >= task.RAM && watts >= task.Watts {
+		return true
+	}
+
+	return false
+}
+
+type BPMaxMinPistonCapping struct {
+	base         //Type embedding to inherit common functions
+	tasksCreated int
+	tasksRunning int
+	tasks        []def.Task
+	metrics      map[string]def.Metric
+	running      map[string]map[string]bool
+	taskMonitor  map[string][]def.Task
+	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 a 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{}
+
+	schedTrace *log.Logger
+}
+
+// New electron scheduler
+func NewBPMaxMinPistonCapping(tasks []def.Task, ignoreWatts bool, schedTracePrefix string) *BPMaxMinPistonCapping {
+	sort.Sort(def.WattsSorter(tasks))
+
+	logFile, err := os.Create("./" + schedTracePrefix + "_schedTrace.log")
+	if err != nil {
+		log.Fatal(err)
+	}
+
+	s := &BPMaxMinPistonCapping{
+		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),
+		totalPower:  make(map[string]float64),
+		RecordPCP:   false,
+		ticker:      time.NewTicker(5 * time.Second),
+		isCapping:   false,
+		schedTrace:  log.New(logFile, "", log.LstdFlags),
+	}
+	return s
+
+}
+
+func (s *BPMaxMinPistonCapping) newTask(offer *mesos.Offer, task def.Task) *mesos.TaskInfo {
+	taskName := fmt.Sprintf("%s-%d", task.Name, *task.Instances)
+	s.tasksCreated++
+
+	// Start recording only when we're creating the first task
+	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)
+	}
+
+	// Add task to list of tasks running on node
+	s.running[offer.GetSlaveId().GoString()][taskName] = true
+
+	// 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
+	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 *BPMaxMinPistonCapping) Disconnected(sched.SchedulerDriver) {
+	// Need to stop the capping process
+	s.ticker.Stop()
+	bpMaxMinPistonCappingMutex.Lock()
+	s.isCapping = false
+	bpMaxMinPistonCappingMutex.Unlock()
+	log.Println("Framework disconnected with master")
+}
+
+// mutex
+var bpMaxMinPistonCappingMutex sync.Mutex
+
+// go routine to cap each node in the cluster at regular intervals of time
+var bpMaxMinPistonCappingCapValues = 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 bpMaxMinPistonCappingPreviousRoundedCapValues = make(map[string]int)
+
+func (s *BPMaxMinPistonCapping) startCapping() {
+	go func() {
+		for {
+			select {
+			case <-s.ticker.C:
+				// Need to cap each node
+				bpMaxMinPistonCappingMutex.Lock()
+				for host, capValue := range bpMaxMinPistonCappingCapValues {
+					roundedCapValue := int(math.Floor(capValue + 0.5))
+					// has the cap value changed
+					if previousRoundedCap, ok := bpMaxMinPistonCappingPreviousRoundedCapValues[host]; ok {
+						if previousRoundedCap != roundedCapValue {
+							if err := rapl.Cap(host, "rapl", roundedCapValue); err != nil {
+								log.Println(err)
+							} else {
+								log.Printf("Capped [%s] at %d", host, int(math.Floor(capValue)))
+							}
+							bpMaxMinPistonCappingPreviousRoundedCapValues[host] = roundedCapValue
+						}
+					} else {
+						if err := rapl.Cap(host, "rapl", roundedCapValue); err != nil {
+							log.Println(err)
+						} else {
+							log.Printf("Capped [%s] at %d", host, int(math.Floor(capValue+0.5)))
+						}
+						bpMaxMinPistonCappingPreviousRoundedCapValues[host] = roundedCapValue
+					}
+				}
+				bpMaxMinPistonCappingMutex.Unlock()
+			}
+		}
+	}()
+
+}
+
+// Stop the capping
+func (s *BPMaxMinPistonCapping) stopCapping() {
+	if s.isCapping {
+		log.Println("Stopping the capping.")
+		s.ticker.Stop()
+		bpMaxMinPistonCappingMutex.Lock()
+		s.isCapping = false
+		bpMaxMinPistonCappingMutex.Unlock()
+	}
+}
+
+// Determine if the remaining sapce inside of the offer is enough for
+// the task we need to create. If it is, create a TaskInfo and return it.
+func (s *BPMaxMinPistonCapping) CheckFit(i int,
+	task def.Task,
+	offer *mesos.Offer,
+	totalCPU *float64,
+	totalRAM *float64,
+	totalWatts *float64,
+	partialLoad *float64) (bool, *mesos.TaskInfo) {
+
+	offerCPU, offerRAM, offerWatts := OfferAgg(offer)
+
+	// Does the task fit
+	if (s.ignoreWatts || (offerWatts >= (*totalWatts + task.Watts))) &&
+		(offerCPU >= (*totalCPU + task.CPU)) &&
+		(offerRAM >= (*totalRAM + task.RAM)) {
+
+		// Start piston capping if haven't started yet
+		if !s.isCapping {
+			s.isCapping = true
+			s.startCapping()
+		}
+
+		*totalWatts += task.Watts
+		*totalCPU += task.CPU
+		*totalRAM += task.RAM
+		log.Println("Co-Located with: ")
+		coLocated(s.running[offer.GetSlaveId().GoString()])
+
+		taskToSchedule := s.newTask(offer, task)
+
+		fmt.Println("Inst: ", *task.Instances)
+		s.schedTrace.Print(offer.GetHostname() + ":" + taskToSchedule.GetTaskId().GetValue())
+		*task.Instances--
+		*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)
+			}
+		}
+
+		return true, taskToSchedule
+	}
+
+	return false, nil
+}
+
+func (s *BPMaxMinPistonCapping) ResourceOffers(driver sched.SchedulerDriver, offers []*mesos.Offer) {
+	log.Printf("Received %d resource offers", len(offers))
+
+	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:
+		}
+
+		tasks := []*mesos.TaskInfo{}
+
+		offerTaken := 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 the capValue for this host using partialLoad and then launch the fit tasks.
+		partialLoad := 0.0
+
+		// Assumes s.tasks is ordered in non-decreasing median max peak order
+
+		// Attempt to schedule a single instance of the heaviest workload available first
+		// Start from the back until one fits
+		for i := len(s.tasks) - 1; i >= 0; i-- {
+
+			task := s.tasks[i]
+			// Check host if it exists
+			if task.Host != "" {
+				// Don't take offer if it doesn't match our task's host requirement
+				if !strings.HasPrefix(*offer.Hostname, task.Host) {
+					continue
+				}
+			}
+
+			// TODO: Fix this so index doesn't need to be passed
+			taken, taskToSchedule := s.CheckFit(i, task, offer, &totalCPU, &totalRAM, &totalWatts, &partialLoad)
+
+			if taken {
+				offerTaken = true
+				tasks = append(tasks, taskToSchedule)
+				break
+			}
+		}
+
+		// Pack the rest of the offer with the smallest tasks
+		for i, task := range s.tasks {
+
+			// Check host if it exists
+			if task.Host != "" {
+				// Don't take offer if it doesn't match our task's host requirement
+				if !strings.HasPrefix(*offer.Hostname, task.Host) {
+					continue
+				}
+			}
+
+			for *task.Instances > 0 {
+				// TODO: Fix this so index doesn't need to be passed
+				taken, taskToSchedule := s.CheckFit(i, task, offer, &totalCPU, &totalRAM, &totalWatts, &partialLoad)
+
+				if taken {
+					offerTaken = true
+					tasks = append(tasks, taskToSchedule)
+				} else {
+					break // Continue on to next task
+				}
+			}
+		}
+
+		if offerTaken {
+			// Updating the cap value for offer.Hostname
+			bpMaxMinPistonCappingMutex.Lock()
+			bpMaxMinPistonCappingCapValues[*offer.Hostname] += partialLoad
+			bpMaxMinPistonCappingMutex.Unlock()
+			log.Printf("Starting on [%s]\n", offer.GetHostname())
+			driver.LaunchTasks([]*mesos.OfferID{offer.Id}, tasks, defaultFilter)
+		} else {
+
+			// If there was no match for the task
+			fmt.Println("There is not enough resources to launch a 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 *BPMaxMinPistonCapping) 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 *BPMaxMinPistonCapping) StatusUpdate(driver sched.SchedulerDriver, status *mesos.TaskStatus) {
+	log.Printf("Received task status [%s] for task [%s]", NameFor(status.State), *status.TaskId.Value)
+
+	if *status.State == mesos.TaskState_TASK_RUNNING {
+		bpMaxMinPistonCappingMutex.Lock()
+		s.tasksRunning++
+		bpMaxMinPistonCappingMutex.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
+		bpMaxMinPistonCappingMutex.Lock()
+		bpMaxMinPistonCappingCapValues[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(bpMaxMinPistonCappingCapValues[hostOfFinishedTask]+0.5)) == 0 {
+			bpMaxMinPistonCappingCapValues[hostOfFinishedTask] = 100
+		}
+		s.tasksRunning--
+		bpMaxMinPistonCappingMutex.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)
+
+}

schedulers/bpMaxMinProacCC.go

@@ -0,0 +1,447 @@
+package schedulers
+
+import (
+	"bitbucket.org/sunybingcloud/electron/constants"
+	"bitbucket.org/sunybingcloud/electron/def"
+	"bitbucket.org/sunybingcloud/electron/pcp"
+	"bitbucket.org/sunybingcloud/electron/rapl"
+	"fmt"
+	"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"
+	"os"
+	"sort"
+	"strings"
+	"sync"
+	"time"
+)
+
+// Decides if to take an offer or not
+func (s *BPMaxMinProacCC) takeOffer(offer *mesos.Offer, task def.Task) bool {
+	cpus, mem, watts := OfferAgg(offer)
+
+	//TODO: Insert watts calculation here instead of taking them as a parameter
+
+	if cpus >= task.CPU && mem >= task.RAM && watts >= task.Watts {
+		return true
+	}
+
+	return false
+}
+
+type BPMaxMinProacCC struct {
+	base           // Type embedding to inherit common functions
+	tasksCreated   int
+	tasksRunning   int
+	tasks          []def.Task
+	metrics        map[string]def.Metric
+	running        map[string]map[string]bool
+	taskMonitor    map[string][]def.Task
+	availablePower map[string]float64
+	totalPower     map[string]float64
+	ignoreWatts    bool
+	capper         *pcp.ClusterwideCapper
+	ticker         *time.Ticker
+	recapTicker    *time.Ticker
+	isCapping      bool // indicate whether we are currently performing cluster-wide capping.
+	isRecapping    bool // indicate whether we are currently performing cluster-wide recapping.
+
+	// First set of PCP values are garbage values, signal to logger to start recording when we're
+	// about to schedule a 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{}
+
+	schedTrace *log.Logger
+}
+
+// New electron scheduler
+func NewBPMaxMinProacCC(tasks []def.Task, ignoreWatts bool, schedTracePrefix string) *BPMaxMinProacCC {
+	sort.Sort(def.WattsSorter(tasks))
+
+	logFile, err := os.Create("./" + schedTracePrefix + "_schedTrace.log")
+	if err != nil {
+		log.Fatal(err)
+	}
+
+	s := &BPMaxMinProacCC{
+		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),
+		availablePower: make(map[string]float64),
+		totalPower:     make(map[string]float64),
+		RecordPCP:      false,
+		capper:         pcp.GetClusterwideCapperInstance(),
+		ticker:         time.NewTicker(10 * time.Second),
+		recapTicker:    time.NewTicker(20 * time.Second),
+		isCapping:      false,
+		isRecapping:    false,
+		schedTrace:     log.New(logFile, "", log.LstdFlags),
+	}
+	return s
+}
+
+// mutex
+var bpMaxMinProacCCMutex sync.Mutex
+
+func (s *BPMaxMinProacCC) 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 the list of tasks running on the node.
+	s.running[offer.GetSlaveId().GoString()][taskName] = true
+	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
+			},
+		},
+	}
+}
+
+// go routine to cap the entire cluster in regular intervals of time.
+var bpMaxMinProacCCCapValue = 0.0 // initial value to indicate that we haven't capped the cluster yet.
+var bpMaxMinProacCCNewCapValue = 0.0 // newly computed cap value
+func (s *BPMaxMinProacCC) startCapping() {
+	go func() {
+		for {
+			select {
+			case <-s.ticker.C:
+				// Need to cap the cluster only if new cap value different from old cap value.
+				// This way we don't unnecessarily cap the cluster.
+				bpMaxMinProacCCMutex.Lock()
+				if s.isCapping {
+					if int(math.Floor(bpMaxMinProacCCNewCapValue+0.5)) != int(math.Floor(bpMaxMinProacCCCapValue+0.5)) {
+						// updating cap value
+						bpMaxMinProacCCCapValue = bpMaxMinProacCCNewCapValue
+						if bpMaxMinProacCCCapValue > 0.0 {
+							for _, host := range constants.Hosts {
+								// Rounding cap value to nearest int
+								if err := rapl.Cap(host, "rapl", int(math.Floor(bpMaxMinProacCCCapValue+0.5))); err != nil {
+									log.Println(err)
+								}
+							}
+							log.Printf("Capped the cluster to %d", int(math.Floor(bpMaxMinProacCCCapValue+0.5)))
+						}
+					}
+				}
+				bpMaxMinProacCCMutex.Unlock()
+			}
+		}
+	}()
+}
+
+// go routine to recap the entire cluster in regular intervals of time.
+var bpMaxMinProacCCRecapValue = 0.0 // The cluster-wide cap value when recapping.
+func (s *BPMaxMinProacCC) startRecapping() {
+	go func() {
+		for {
+			select {
+			case <-s.recapTicker.C:
+				bpMaxMinProacCCMutex.Lock()
+				// If stopped performing cluster-wide capping, then we need to recap.
+				if s.isRecapping && bpMaxMinProacCCRecapValue > 0.0 {
+					for _, host := range constants.Hosts {
+						// Rounding the recap value to the nearest int
+						if err := rapl.Cap(host, "rapl", int(math.Floor(bpMaxMinProacCCRecapValue+0.5))); err != nil {
+							log.Println(err)
+						}
+					}
+					log.Printf("Capped the cluster to %d", int(math.Floor(bpMaxMinProacCCRecapValue+0.5)))
+				}
+				// Setting the recapping to false
+				s.isRecapping = false
+				bpMaxMinProacCCMutex.Unlock()
+			}
+		}
+	}()
+
+}
+
+// Stop cluster-wide capping
+func (s *BPMaxMinProacCC) stopCapping() {
+	if s.isCapping {
+		log.Println("Stopping the cluster-wide capping.")
+		s.ticker.Stop()
+		bpMaxMinProacCCMutex.Lock()
+		s.isCapping = false
+		s.isRecapping = true
+		bpMaxMinProacCCMutex.Unlock()
+	}
+}
+
+// Stop the cluster-wide recapping
+func (s *BPMaxMinProacCC) stopRecapping() {
+	// If not capping, then definitely recapping.
+	if !s.isCapping && s.isRecapping {
+		log.Println("Stopping the cluster-wide re-capping.")
+		s.recapTicker.Stop()
+		bpMaxMinProacCCMutex.Lock()
+		s.isRecapping = false
+		bpMaxMinProacCCMutex.Unlock()
+	}
+}
+
+// Determine if the remaining space inside of the offer is enough for
+// the task we need to create. If it is, create TaskInfo and return it.
+func (s *BPMaxMinProacCC) CheckFit(i int,
+	task def.Task,
+	offer *mesos.Offer,
+	totalCPU *float64,
+	totalRAM *float64,
+	totalWatts *float64) (bool, *mesos.TaskInfo) {
+
+	offerCPU, offerRAM, offerWatts := OfferAgg(offer)
+
+	// Does the task fit
+	if (s.ignoreWatts || (offerWatts >= (*totalWatts + task.Watts))) &&
+		(offerCPU >= (*totalCPU + task.CPU)) &&
+		(offerRAM >= (*totalRAM + task.RAM)) {
+
+		// Capping the cluster if haven't yet started
+		if !s.isCapping {
+			bpMaxMinProacCCMutex.Lock()
+			s.isCapping = true
+			bpMaxMinProacCCMutex.Unlock()
+			s.startCapping()
+		}
+
+		tempCap, err := s.capper.FCFSDeterminedCap(s.totalPower, &task)
+		if err == nil {
+			bpMaxMinProacCCMutex.Lock()
+			bpMaxMinProacCCNewCapValue = tempCap
+			bpMaxMinProacCCMutex.Unlock()
+		} else {
+			log.Println("Failed to determine new cluster-wide cap:")
+			log.Println(err)
+		}
+
+		*totalWatts += task.Watts
+		*totalCPU += task.CPU
+		*totalRAM += task.RAM
+		log.Println("Co-Located with: ")
+		coLocated(s.running[offer.GetSlaveId().GoString()])
+
+		taskToSchedule := s.newTask(offer, task)
+
+		fmt.Println("Inst: ", *task.Instances)
+		s.schedTrace.Print(offer.GetHostname() + ":" + taskToSchedule.GetTaskId().GetValue())
+		*task.Instances--
+
+		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")
+				// Need to stop the cluster wide capping
+				s.stopCapping()
+				s.startRecapping() // Load changes after every task finishes and hence, we need to change the capping of the cluster.
+				close(s.Shutdown)
+			}
+		}
+
+		return true, taskToSchedule
+	}
+
+	return false, nil
+
+}
+
+func (s *BPMaxMinProacCC) ResourceOffers(driver sched.SchedulerDriver, offers []*mesos.Offer) {
+	log.Printf("Received %d resource offers", len(offers))
+
+	// retrieving the available power for all the hosts in the offers.
+	for _, offer := range offers {
+		_, _, offerWatts := OfferAgg(offer)
+		s.availablePower[*offer.Hostname] = offerWatts
+		// setting total power if the first time
+		if _, ok := s.totalPower[*offer.Hostname]; !ok {
+			s.totalPower[*offer.Hostname] = offerWatts
+		}
+	}
+
+	for host, tpower := range s.totalPower {
+		log.Printf("TotalPower[%s] = %f", host, tpower)
+	}
+
+	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:
+		}
+
+		tasks := []*mesos.TaskInfo{}
+
+		offerTaken := false
+		totalWatts := 0.0
+		totalCPU := 0.0
+		totalRAM := 0.0
+
+		// Assumes s.tasks is ordered in non-decreasing median max peak order
+
+		// Attempt to schedule a single instance of the heaviest workload available first
+		// Start from the back until one fits
+		for i := len(s.tasks) - 1; i >= 0; i-- {
+
+			task := s.tasks[i]
+			// Check host if it exists
+			if task.Host != "" {
+				// Don't take offer if it doesn't match our task's host requirement
+				if !strings.HasPrefix(*offer.Hostname, task.Host) {
+					continue
+				}
+			}
+
+			// TODO: Fix this so index doesn't need to be passed
+			taken, taskToSchedule := s.CheckFit(i, task, offer, &totalCPU, &totalRAM, &totalWatts)
+
+			if taken {
+				offerTaken = true
+				tasks = append(tasks, taskToSchedule)
+				break
+			}
+		}
+
+		// Pack the rest of the offer with the smallest tasks
+		for i, task := range s.tasks {
+
+			// Check host if it exists
+			if task.Host != "" {
+				// Don't take offer if it doesn't match our task's host requirement
+				if !strings.HasPrefix(*offer.Hostname, task.Host) {
+					continue
+				}
+			}
+
+			for *task.Instances > 0 {
+				// TODO: Fix this so index doesn't need to be passed
+				taken, taskToSchedule := s.CheckFit(i, task, offer, &totalCPU, &totalRAM, &totalWatts)
+
+				if taken {
+					offerTaken = true
+					tasks = append(tasks, taskToSchedule)
+				} else {
+					break // Continue on to next task
+				}
+			}
+		}
+
+		if offerTaken {
+			log.Printf("Starting on [%s]\n", offer.GetHostname())
+			driver.LaunchTasks([]*mesos.OfferID{offer.Id}, tasks, defaultFilter)
+		} else {
+
+			// If there was no match for the task
+			fmt.Println("There is not enough resources to launch a task:")
+			cpus, mem, watts := OfferAgg(offer)
+
+			log.Printf("<CPU: %f, RAM: %f, Watts: %f>\n", cpus, mem, watts)
+			driver.DeclineOffer(offer.Id, defaultFilter)
+		}
+	}
+}
+
+func (s *BPMaxMinProacCC) StatusUpdate(driver sched.SchedulerDriver, status *mesos.TaskStatus) {
+	log.Printf("Received task status [%s] for task [%s]", NameFor(status.State), *status.TaskId.Value)
+
+	if *status.State == mesos.TaskState_TASK_RUNNING {
+		s.tasksRunning++
+	} else if IsTerminal(status.State) {
+		delete(s.running[status.GetSlaveId().GoString()], *status.TaskId.Value)
+		// Need to remove the task from the window
+		s.capper.TaskFinished(*status.TaskId.Value)
+		// Determining the new cluster wide recap value
+		tempCap, err := s.capper.NaiveRecap(s.totalPower, s.taskMonitor, *status.TaskId.Value)
+		//tempCap, err := s.capper.CleverRecap(s.totalPower, s.taskMonitor, *status.TaskId.Value)
+		if err == nil {
+			// If new determined recap value is different from the current recap value, then we need to recap.
+			if int(math.Floor(tempCap+0.5)) != int(math.Floor(bpMaxMinProacCCRecapValue+0.5)) {
+				bpMaxMinProacCCRecapValue = tempCap
+				bpMaxMinProacCCMutex.Lock()
+				s.isRecapping = true
+				bpMaxMinProacCCMutex.Unlock()
+				log.Printf("Determined re-cap value: %f\n", bpMaxMinProacCCRecapValue)
+			} else {
+				bpMaxMinProacCCMutex.Lock()
+				s.isRecapping = false
+				bpMaxMinProacCCMutex.Unlock()
+			}
+		} else {
+			log.Println(err)
+		}
+
+		s.tasksRunning--
+		if s.tasksRunning == 0 {
+			select {
+			case <-s.Shutdown:
+				// Need to stop the cluster-wide recapping
+				s.stopRecapping()
+				close(s.Done)
+			default:
+			}
+		}
+	}
+	log.Printf("DONE: Task status [%s] for task [%s]", NameFor(status.State), *status.TaskId.Value)
+
+}

schedulers/bpswClassMapWattsPistonCapping.go

@@ -104,6 +104,9 @@ func (s *BPSWClassMapWattsPistonCapping) newTask(offer *mesos.Offer, task def.Ta
 		s.running[offer.GetSlaveId().GoString()] = make(map[string]bool)
 	}
 
+	// Add task to list of tasks running on node
+	s.running[offer.GetSlaveId().GoString()][taskName] = true
+
 	// 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))
@@ -155,7 +158,7 @@ func (s *BPSWClassMapWattsPistonCapping) Disconnected(sched.SchedulerDriver) {
 // mutex
 var bpswClassMapWattsPistonMutex sync.Mutex
 
-// go routine to cap eahc node in the cluster at regular intervals of time
+// go routine to cap each node in the cluster at regular intervals of time
 var bpswClassMapWattsPistonCapValues = 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)
@@ -268,6 +271,12 @@ func (s *BPSWClassMapWattsPistonCapping) ResourceOffers(driver sched.SchedulerDr
 					(offerCPU >= (totalCPU + task.CPU)) &&
 					(offerRAM >= (totalRAM + task.RAM)) {
 
+					// Start piston capping if haven't started yet
+					if !s.isCapping {
+						s.isCapping = true
+						s.startCapping()
+					}
+
 					fmt.Println("Watts being used: ", task.ClassToWatts[nodeClass])
 					taken = true
 					totalWatts += task.ClassToWatts[nodeClass]

schedulers/bpswClassMapWattsProacCC.go

@@ -165,22 +165,30 @@ func (s *BPSWClassMapWattsProacCC) Disconnected(sched.SchedulerDriver) {
 }
 
 // go routine to cap the entire cluster in regular intervals of time.
-var bpswClassMapWattsCapValue = 0.0 // initial value to indicate that we haven't capped the cluster yet.
+var bpswClassMapWattsProacCCCapValue = 0.0 // initial value to indicate that we haven't capped the cluster yet.
+var bpswClassMapWattsProacCCNewCapValue = 0.0 // newly computed cap value
 func (s *BPSWClassMapWattsProacCC) startCapping() {
 	go func() {
 		for {
 			select {
 			case <-s.ticker.C:
-				// Need to cap the cluster to the bpswClassMapWattsCapValue.
+				// Need to cap the cluster only if new cap value different from old cap value.
+				// This way we don't unnecessarily cap the cluster.
 				bpswClassMapWattsProacCCMutex.Lock()
-				if bpswClassMapWattsCapValue > 0.0 {
+				if s.isCapping {
-					for _, host := range constants.Hosts {
+					if int(math.Floor(bpswClassMapWattsProacCCNewCapValue+0.5)) != int(math.Floor(bpswClassMapWattsProacCCCapValue+0.5)) {
-						// Rounding capValue to nearest int.
+						// updating cap value
-						if err := rapl.Cap(host, "rapl", int(math.Floor(bpswClassMapWattsCapValue+0.5))); err != nil {
+						bpswClassMapWattsProacCCCapValue = bpswClassMapWattsProacCCNewCapValue
-							log.Println(err)
+						if bpswClassMapWattsProacCCCapValue > 0.0 {
+							for _, host := range constants.Hosts {
+								// Rounding cap value to nearest int
+								if err := rapl.Cap(host, "rapl", int(math.Floor(bpswClassMapWattsProacCCCapValue+0.5))); err != nil {
+									log.Println(err)
+								}
+							}
+							log.Printf("Capped the cluster to %d", int(math.Floor(bpswClassMapWattsProacCCCapValue+0.5)))
 						}
 					}
-					log.Printf("Capped the cluster to %d", int(math.Floor(bpswClassMapWattsCapValue+0.5)))
 				}
 				bpswClassMapWattsProacCCMutex.Unlock()
 			}
@@ -189,7 +197,7 @@ func (s *BPSWClassMapWattsProacCC) startCapping() {
 }
 
 // go routine to recap the entire cluster in regular intervals of time.
-var bpswClassMapWattsRecapValue = 0.0 // The cluster-wide cap value when recapping
+var bpswClassMapWattsProacCCRecapValue = 0.0 // The cluster-wide cap value when recapping
 func (s *BPSWClassMapWattsProacCC) startRecapping() {
 	go func() {
 		for {
@@ -197,14 +205,14 @@ func (s *BPSWClassMapWattsProacCC) startRecapping() {
 			case <-s.recapTicker.C:
 				bpswClassMapWattsProacCCMutex.Lock()
 				// If stopped performing cluster wide capping, then we need to recap
-				if s.isRecapping && bpswClassMapWattsRecapValue > 0.0 {
+				if s.isRecapping && bpswClassMapWattsProacCCRecapValue > 0.0 {
 					for _, host := range constants.Hosts {
 						// Rounding capValue to the nearest int
-						if err := rapl.Cap(host, "rapl", int(math.Floor(bpswClassMapWattsRecapValue+0.5))); err != nil {
+						if err := rapl.Cap(host, "rapl", int(math.Floor(bpswClassMapWattsProacCCRecapValue +0.5))); err != nil {
 							log.Println(err)
 						}
 					}
-					log.Printf("Recapping the cluster to %d", int(math.Floor(bpswClassMapWattsRecapValue+0.5)))
+					log.Printf("Recapping the cluster to %d", int(math.Floor(bpswClassMapWattsProacCCRecapValue +0.5)))
 				}
 				// Setting recapping to false
 				s.isRecapping = false
@@ -309,7 +317,7 @@ func (s *BPSWClassMapWattsProacCC) ResourceOffers(driver sched.SchedulerDriver,
 					tempCap, err := s.capper.FCFSDeterminedCap(s.totalPower, &task)
 					if err == nil {
 						bpswClassMapWattsProacCCMutex.Lock()
-						bpswClassMapWattsCapValue = tempCap
+						bpswClassMapWattsProacCCNewCapValue = tempCap
 						bpswClassMapWattsProacCCMutex.Unlock()
 					} else {
 						log.Println("Failed to determine new cluster-wide cap:")
@@ -370,16 +378,16 @@ func (s *BPSWClassMapWattsProacCC) StatusUpdate(driver sched.SchedulerDriver, st
 		// Need to remove the task from the window
 		s.capper.TaskFinished(*status.TaskId.Value)
 		// Determining the new cluster wide recap value
+		//tempCap, err := s.capper.NaiveRecap(s.totalPower, s.taskMonitor, *status.TaskId.Value)
 		tempCap, err := s.capper.CleverRecap(s.totalPower, s.taskMonitor, *status.TaskId.Value)
-		//tempCap, err := s.capper.CleverRecap(s.totalPower, s.taskMonitor, *status.TaskId.Value)
 		if err == nil {
 			// If new determined cap value is different from the current recap value, then we need to recap
-			if int(math.Floor(tempCap+0.5)) != int(math.Floor(bpswClassMapWattsRecapValue+0.5)) {
+			if int(math.Floor(tempCap+0.5)) != int(math.Floor(bpswClassMapWattsProacCCRecapValue +0.5)) {
-				bpswClassMapWattsRecapValue = tempCap
+				bpswClassMapWattsProacCCRecapValue = tempCap
 				bpswClassMapWattsProacCCMutex.Lock()
 				s.isRecapping = true
 				bpswClassMapWattsProacCCMutex.Unlock()
-				log.Printf("Determined re-cap value: %f\n", bpswClassMapWattsRecapValue)
+				log.Printf("Determined re-cap value: %f\n", bpswClassMapWattsProacCCRecapValue)
 			} else {
 				bpswClassMapWattsProacCCMutex.Lock()
 				s.isRecapping = false

schedulers/proactiveclusterwidecappingfcfs.go

@@ -348,7 +348,7 @@ func (s *ProactiveClusterwideCapFCFS) StatusUpdate(driver sched.SchedulerDriver,
 		// Need to remove the task from the window of tasks.
 		s.capper.TaskFinished(*status.TaskId.Value)
 		// Determining the new cluster wide cap.
-		//tempCap, err := s.capper.Recap(s.totalPower, s.taskMonitor, *status.TaskId.Value)
+		//tempCap, err := s.capper.NaiveRecap(s.totalPower, s.taskMonitor, *status.TaskId.Value)
 		tempCap, err := s.capper.CleverRecap(s.totalPower, s.taskMonitor, *status.TaskId.Value)
 		if err == nil {
 			// if new determined cap value is different from the current recap value then we need to recap.

schedulers/proactiveclusterwidecappingranked.go

@@ -383,7 +383,7 @@ func (s *ProactiveClusterwideCapRanked) StatusUpdate(driver sched.SchedulerDrive
 			// Need to remove the task from the window
 			s.capper.TaskFinished(*status.TaskId.Value)
 			// Determining the new cluster wide cap.
-			//tempCap, err := s.capper.Recap(s.totalPower, s.taskMonitor, *status.TaskId.Value)
+			//tempCap, err := s.capper.NaiveRecap(s.totalPower, s.taskMonitor, *status.TaskId.Value)
 			tempCap, err := s.capper.CleverRecap(s.totalPower, s.taskMonitor, *status.TaskId.Value)
 
 			if err == nil {