Computer Science/162/proj1

KThread.join()

Implementation

New state variables

KThread has a new state variable joinedOnMe, a ThreadQueue.

Implementation details

• When threadA calls threadB.join(), threadB adds it to its internal joinedOnMe queue and then puts it to sleep. In threadB's finish() method, threadB calls nextThread() on its joinedOnMe queue before returning.
• join() also checks that this thread is not equal to the current thread and that joinedOnMe is not already occupied by another thread with a boolean flag, which determines whether join() has already been called on this thread.

Testing

• threadA prints out a few statements, calls threadB.join(), then prints out a few more statements. threadB also prints out a series of statements. We then verify that threadB executes contiguously to completion before threadA resumes its execution, as evidenced by the sequence of printed statements.
• Have a thread attempt to join() to itself, call join() on a second thread multiple times, and attempt to call join() on a third finished thread (all separately). These should all return immediately as this is the expected behavior.
• Test a chain of threads. Thread C will attempt to join Thread B. Thread B will attempt to join Thread A. Thread A forks last. Verify that A executes before B, and B executes before C.

Pseudocode

join() {
    Disable interrupts;
    If (CurrentThread == self or isJoined) or (status is Finished) {
        Re-enable interrupts;
        Return; // conditions for join not satisfied
    } else {
        add CurrentThread to joinedOnMe queue;
        isJoined = true;
        Sleep the currentThread;
    }
    Re-enable interrupts;
}

finish() {
    // Interrupts have been disabled.
    ...(existing code)...
    Ready the thread on the joinedOnMe queue;
    Sleep (existing);
}

Condition2.java

Implementation

New state variables

Condition2 has a new state variable waitQueue, a ThreadQueue with transferPriority flag set to false.

Implementation details

• In the sleep() method, we disable interrupts, the current thread is added to the waitQueue of threads, the corresponding lock is released, and the current thread is put to sleep. When it wakes up, the thread attempts to reacquire the corresponding lock and interrupts are re-enabled.
• The wake() method disables interrupts and retrieves the next thread from waitQueue. If the thread returned is not null, that thread is readied. Finally, the interrupts are re-enabled.
• The wakeAll() method also disables the interrupts and stores the next thread from waitQueue as a KThread variable, nextThread. While nextThread is not null, the method will ready that thread and set nextThread to the next thread from waitQueue. This procedure continues until there are no more threads in waitQueue.

Testing

• Condition2.java was tested using a class called CondThread, which implements Runnable. The class is instantiated with an integer iterations, a lock, a Condition2 variable to call sleep() on, and a Condition2 variable to call wake() on. When a CondThread runs, it acquires the lock, then enters a for-loop of 6 iterations. On each iteration, the thread will print out a debug statement noting that the current thread is on the ith iteration. When the loop iteration number is equal to iterations, threads can take one of three actions:
  1. If the thread's ID is '0', then that thread is the waking thread and calls wake() on its second condition variable.
  2. If the thread's ID is '0-all', it calls wakeAll() on its second condition variable
  3. Otherwise, it goes to sleep on its first condition variable. Using the output, we verify that the threads all run in the correct order and all execute to completion.
• Sleep several threads on the condition variable. Upon waking, these threads should print a unique statement and perform a wake() on the condition variable. Have one thread perform an initial wake() on the condition variable and verify that all threads are executed.
• Sleep several threads on several various condition variables. Verify via console print statements that the proper threads wake up at the correct times according to which condition variable they were sleeping on.
• Sleep several threads on the condition variable, then have one final thread wake them all with wakeAll(). Verify that all threads have woken up via console print statements.
• Sleep several threads on several various condition variables, then have one thread call wakeAll() on one of the condition variables. Verify via console print statements that all the threads put to sleep on that condition variable wake up and that only those threads wake up.

Pseudocode

sleep() { 
    Disable Interrupts;
    Add current thread to wait queue;
    Release the lock;
    Put the current thread to sleep;
    Acquire the lock;
    Re-enable Interrupts;
}

wake() {
    AssertTrue (lock held by current thread);
    Disable interrupts;
    If there is a thread on the wait queue:
        Remove the first thread and wake it;
    Re-enable interrupts; 
}

wakeAll() {
    Disable interrupts;
    While there are threads on the wait queue:
        Remove the first thread and wake it;
    Re-enable interrupts; 
}

WaitUntil()

Implementation

New state variables

Alarm has a new instance variable, waitingThreads, which is a Java PriorityQueue of waiting threads with the target wake time as their priority. It also contains a new inner class named waitingThread, which contains a reference to a KThread and its associated wakeTime. The PriorityQueue waitingThreads will be populated with instances of this inner class.

Implementation details

• waitUntil() creates a new instance of waitingThread, which will associate the currentThread with the give time argument. It will then add this waitingThread instance to the PriorityQueue and put the current thread to sleep. This method is atomic.
• timerInterrupt() peek at the first WaitingThread in the waitQueue and check if its associated wake time is less than or equal to the current time. If it is, this method will pop the WaitingThread off the waitQueue and wake the associated thread. This process is repeated until the wake time of the first WaitingThread is greater than the current time.

Testing

• Have the timer interrupt print a debug statement, and have timed threads print the times they go to sleep and when they have woken up. Verify that they wake up relatively close to their expected exit time.
• Make sure threads called with illegal times return immediately. Test threads that go to sleep chronologically in order (1, 500, 1050) as well as threads that go to sleep with the same times (50, 50) and threads that go to sleep in reverse order (500, 250). Verify they all wake up when they are supposed to.

Pseudocode

waitUntil(time){
    Disable interrupts;
    Create a new waitingThread;
    Add the waiting thread to the priority queue;
    Sleep the current thread;
    Re-enable interrupts;
}

timerInterrupt(){
    AssertTrue (interrupts have already been disabled);
    For all waitingThreads that have exceeded their associated wait time;
    Wake their associated threads and remove from queue;
}

Communicator

Implementation

New state variables

Lock lock = new Lock()
int activeSpeakers = 0
int waitingSpeakers = 0
int activeListeners = 0
int waitingListeners = 0
Condition speakers
Condition listeners
Condition return

Implementation details

• The first lone speaker or listener will be counted as active, or in the process of exchanging a message and returning, and will sleep on the return condition variable until its counterpart wakes it up so that they can both return.
• A second thread performing the same action as a currently active thread will be counted as waiting, and be put to sleep on its respective condition variable. Otherwise, it will check if there is an active thread of its counterpart action waiting on the return condition variable. If there isn't, it will attempt to wake waiting threads of its counterpart action prior to going to sleep on the return condition variable. If there is a counterpart active thread, it will wake it up and they both will return. Prior to returning, the counterpart action will also attempt to wake waiting threads of its type.
• Any interjecting threads that execute in between an exchange of message will be stopped by the active counters, which do not decrement until both counterparts in the exchange have returned.

Testing

Our original solution exhibited non-deterministic behavior, so after rewriting it, we decided to stress it exceptionally to make sure that it was working correctly. We tested our communicator in three main ways.

First, we set up a manual sequence of speak() and listen() actions on different threads and executed them in a particular order, verifying that the resulting sequence of messages was correct, monitored via print statements to the console

Second, we set off a random number of speakers, followed by a random number of listeners, and verified that the limiting resource was completely used up, i.e. that a matching number of speakers and listeners returned and that this was equal to the smaller of numbers of created speakers/listeners.

Finally, we repeated the same procedure, but intermingled the creation of speakers and listeners via .fork(), such that listeners would start listening before all the speakers were queued up.

The latter two tests were run with up to 500 threads of speakers and listeners each (with a temporary override on the number of max threads in Nachos) and the number of listen and speak operations was analyzed via script. The speakers and listeners would print statements while executing code, which allowed us to perform this analysis.

To be able to perform these tests, we created a number of helper classes which implement Runnable. MassSpeaker and MassListener are designed to be run by a single thread each. These will iterate until a given limit, and on each iteration, there is a 50% chance that a speak (or listen) is called. After each iteration, the thread yields to the opposite thread to do the same. Debug statements will display what threads are doing at each iteration and how messages are being exchanged. With this we can generate large amounts of calls with randomized orders between two threads, and will be able to verify all speaks are correctly received by a listen.

A second set of runnables, MassTSpeaker and MassTListener, are designed to fork off several threads themselves, with each of these forked threads performing a single speak or listen. These forked threads are also executed with 50% chance on each iteration to provide random ordering. We can also verify if all threads are correctly paired off via print statements to console.

Pseudocode

speak(int word) { Acquire the lock; while (There is an active speaker) { WS++; Sleep as a waiting speaker; WS--; } AS++; Set the word; if (There is an active listener) { Wake the active listener; Release the lock; Return; } else { if (There are waiting listeners) { Wake a waiting listener; } Sleep as an active speaker; AS--; AL--; if (There are waiting speakers) { Wake a waiting speaker; } Release the lock; Return; } }

listen() { Acquire the lock; while (There is an active listener) { WL++; Sleep as a waiting listener; WL--; } AL++; if (There is an active speaker) { Wake an active speaker; Store the word; Release the lock; Return the word; } else { if (There are waiting speakers) { Wake a waiting speaker; } Sleep as an active listener; AL--; AS--; if (There are waiting listeners) { Wake a waiting listener; } Store the word; Release the lock; Return the word; } }

Priority Scheduler

Implementation

New state variables

• PriorityQueue – ThreadState holder, ArrayList<ThreadState> waitQueue, int effective, boolean dirty
  
• ThreadState – int effective, boolean dirty, HashSet<PriorityQueue> myResources, WaitingOn
  

PriorityQueue.holder - this ThreadState corresponds to the holder of the resource signified by the PriorityQueue.waitQueue - an ArrayList of ThreadStates waiting on this resource. Unsorted. PriorityQueue.dirty - set to true when a new thread is added to the queue, or any of the queues in the waitQueue flag themselves as dirty. PriorityQueue.effective - the cached highest of the effective priorities in the waitQueue. This value is invalidated while dirty is true.

ThreadState.myResources - collection of PriorityQueues that signify the Locks or other resources that this thread currently holds. ThreadState.waitingOn - collection of PriorityQueues corresponding to resources that this thread has attempted to acquire but could not. ThreadState.effective - the cached effective priority of this thread. this value is invalidated when dirty is true ThreadState.dirty - set to true when this thread's priority is changed, or when one of the queues in myResources flags itself as dirty.

Implementation details

• PriorityQueue.nextThread()
• This method retrieves and removes the highest priority thread off the Priority-Time ArrayList. It then flags the previous holder of this resource as dirty, and removes the queue from that holder's resource list. It then sets the retrieved thread to this queue's new holder, and flags that thread as dirty as well.
• PriorityQueue.acquire(KThread thread)
• Sets the holder of this queue to the specified thread, bypassing the queue. Resets previous holder and sets dirty flags as in nextThread().
• PriorityQueue.pickNextThread()
• Simply retrieves the highest priority thread off this queue.
• PriorityQueue.setDirty()
• Set this queue's dirty flag, and calls setDirty on the current holder of this thread.
• PriorityQueue.getEffectivePriority()
• If this queue is "dirty", returns the maximum of each of the ThreadStates.getEffectivePriority() in this queue. Those calls in turn become mutually recursive when they call getEffectivePriority on the PriorityQueues in their myResrouces.
• ThreadState.setPriority()
• This method will change the actual priority of the thread associated with the ThreadState. It then does setDirty() on this thread
• ThreadState.setDirty()
• Sets the dirty flag on this thread, then on each of the PriorityQueues that the thread is waiting for, does setDirty on them. Mutually recursive.
• ThreadState.getEffectivePriority
• Like the analogue of this function in PriorityQueue, returns the priority if this thread is not dirty, otherwise recalculates by returning the max of the effective priorities of the PriorityQueues in myResources.

Testing

• Instantiate new threads and set their priorities in decreasing order. Have the threads state their priorities as they execute and verify that they were run in decreasing order according to their priorities.
• Verify donation works by creating a high priority thread and joining it to a low priority thread with a high priority thread already queued.

Pseudocode

PriorityQueue

   	public void waitForAccess(KThread thread)
   	    add this thread to my waitQueue
           thread.waitForAccess(this)
       
   	public void acquire(KThread thread) 
   	    if I have a holder and I transfer priority, remove myself from the holder's resource list
   	    thread.acquire(this)
       
   	public KThread nextThread() 
   	    if I have a holder and I transfer priority, remove myself from the holder's resource list
   	    if waitQueue is empty, return null
           ThreadState firstThread = pickNextThread();
           remove firstThread from waitQueue
           firstThread.acquire(this);
   	    return firstThread
   	}
   	
   	public int getEffectivePriority()
   	    if I do not transfer priority, return minimum priority
   	    if (dirty)
   	        effective = minimum priority;
   	        for each ThreadState t in waitQueue
   	            effective = MAX(effective, t.getEffectivePriority())
   	        dirty = false;
   	    return effective;
   	
       public void setDirty()
           if I do not transfer priority, there is no need to recurse, return
           dirty = true;
           if I have a holder, holder.setDirty()
       
   	protected ThreadState pickNextThread() 
           ThreadState ret = null
           for each ThreadState ts in waitQueue
              if ret is null OR ts has higher priority/time ranking than ret
                 set ret to ts
   	    return ret;

ThreadState {

   	public int getPriority()
   	    return non-donated priority.
   
   	public int getEffectivePriority()
   	    if (dirty) {
               effective = non-donated priority
   	        for each PriorityQueue pq that I am currently holding
   	          effective = MAX(effective, pq.getEffectivePriority)
   	    }
   	    return effective;

   	public void setPriority(int priority) {
   	    set non-donated priority to the argument
   	    setDirty();
   	}
    
       public void setDirty() {
           if already dirty return
           dirty = true;
           for each of the PriorityQueues pq I am waiting on,
             pq.setDirty
       }
   	
   	public void waitForAccess(PriorityQueue waitQueue) {
   	    add the waitQueue to my waitingOn list
   	    if the waitQueue was previously in myResources, remove it and set its holder to null.
           if waitQueue has a holder, set the queue to dirty to signify possible change in the queue's effective priority
   	
   	public void acquire(PriorityQueue waitQueue) {
   	    add waitQueue to myResources list
   	    if waitQueue was in my waitingOn list, remove it
   	    setWaitQueue's holder to me
   	    setDirty();
   	}

Boat.java

Implementation

New state variables (all shared)

lock = new Lock()
boatIsland = Island.A
boatPilot = null
boatDeparting = false
boatLookingForChild = false
childCounter = 0
adultCounterA = 0
adultCounterB = 0
sleepingChildren = 0
waiters = 0

pipe = new Communicator()

Condition variables, all on lock: childrenWaitingOnA, adultsWaitingOnA, childrenWaitingOnB, adultsWaitingOnB, waitingToRide, waitingToPilot

In addition, each thread has a local variable to keep track of which island the person is currently on.

Basic idea

If there are no children on Molokai or if there are no adults left on Oahu, two children from Oahu will pilot over and one will return to Molokai. If there is a child on Molokai, an adult will pilot over to Molokai and the child will bring the boat back to Oahu. This process, carried to completion eventually results in all of the adults and children on Molokai. Each thread will attempt to acquire the lock (which essentially represents control over the boat). If the thread can perform one of the tasks fitting for the current state of the world, it executes it. Otherwise, it will go to sleep on the condition variable corresponding to its current role and location.

Algorithm details

Each child will begin running on Oahu and try to acquire the lock.

If the child is on Oahu and the boat is available for use (i.e. boatLookingForChild == false and boatPilot == null and boatDeparting == false), it will assume the role of pilot and set the boatLookingForChild variable to true, then going to sleep on waitingToPilot. Once a passenger child is found, the pilot will be awoken by the passenger child and pilot them both to Molokai, while the passenger goes to sleep on waitingToRide. He then increments the childCounter, wakes up the passenger child so that he can ride to Molokai, and goes back to sleep.

The passenger, upon arrival at Molokai, increments the childCounter and speaks to begin() the number of children on Molokai. The child passenger then wakes up the pilot and goes to sleep on childrenWaitingOnB. Once the pilot child reawakens, he returns to Oahu, decrementing childCounter before departure.

Upon return to Oahu, if the number of adults on Oahu (according to adultsWaitingOnA) is not 0 and there is still a child left on Molokai to row the boat back after the adult goes across (according to childCounter), the pilot wakes up the adults waiting on A, puts himself to sleep on childrenWaitingOnA, releases the lock after wake, and restarts the looop. Otherwise, he simply releases the lock and restarts the loop.

The adults all begin on Oahu, try to acquire the lock, and each initially increment adultCounterA.

If the adult is on Oahu, the boat is available for use, and there is at least one child on Molokai according to childCounter, the adult pilots to Molokai. Otherwise, he wakes a child on childrenWaitingOnA and goes to sleep on adultsWaitingOnA.

If the adult does pilot across, he increments adultCounterB upon arrival to Molokai, speaks to begin() the value of childCounter, wakes up a child from childrenWaitingOnB so that he can pilot the boat back to Oahu, and then puts himself to sleep on adultsWaitingOnB. The adult is now done and will never be woken again.

If a child is sleeping on Molokai on childrenWaitingOnB and gets woken by an adult, he will decrement childCounter and pilot himself back to Oahu. Upon arrival, he releases the lock and restarts the loop.

For both adults and children, if the boat is busy or inaccessible as determined by the state of the boatPilot, boatDeparting, boatLookingForChild, and boatIsland variables given the person's current location, each will go back to sleep on the appropriate condition variable (i.e. childrenWaitingOnA for child threads on Oahu).

begin() behavior

begin() creates the required number of child and adult threads and then forks them. Then, it begins to listen() on the one shared Communicator (pipe). When begin() receives a message, it checks the "spoken" number as well as adultCounterB against the number of spawned threads to see if everyone is across yet.

The reason that it does not check the value of childCounter directly is because by the time it receives the message, it is possible that a child has already piloted back to Oahu, thus decrementing the counter and preventing begin() from knowing the true number of children on Molokai.

Testing

We tested different numbers of both children and adults and verified by hand that the sequence of events, in terms of rowing from one island to the other, was correct.

First, we tried various numbers of children with no adults at all: the base case of two children, then a small but odd number of children, and then large numbers of both odd and even amounts of children.

Then, we tested having small numbers of children and adults together, with a greater number of children than adults, both odd and even-number combinations. Again, we monitored that the rowing pattern was correct by hand. We also watched the number of people who have arrived at Molokai and made sure that these numbers made sense.

Finally, we tested larger numbers of both adults and children and verified that the rowing patterns were correct. We tested both combinations with more adults than children and vice versa. The rowing pattern, as well as number of people on Molokai, was monitored to make sure no rules were being violated.

Design questions

Why is it fortunate that we did not ask you to implement priority donation for semaphores?

Be damned if I know. Anyone?

A student proposes to solve the boats problem by use of a counter, AdultsOnOahu. Since this number isn't known initially, it will be started at zero, and incremented by each adult thread before they do anything else. Is this solution likely to work? Why or why not?

No, because there is no way of enforcing the fact that everyone will increment it before they do anything else.