Selection¶

All pools require a Selector to be mixed-in. This trait controls how and how many actors in the pool will receive the incoming message. Define selectionCount to some positive number greater than one to route to multiple actors. Currently two are provided:

• SmallestMailboxSelector - Using the exact same logic as the iterator of the same name, the pooled actor with the fewest number of pending messages will be chosen.
• RoundRobinSelector - Performs a very simple index-based selection, wrapping around the end of the list, very much like the CyclicIterator does.

Capacity¶

As you’d expect, capacity traits determine how the pool is funded with actors. There are two types of strategies that can be employed:

• FixedCapacityStrategy - When you mix this into your actor pool, you define a pool size and when the pool is started, it will have that number of actors within to which messages will be delegated.
• BoundedCapacityStrategy - When you mix this into your actor pool, you define upper and lower bounds, and when the pool is started, it will have the minimum number of actors in place to handle messages. You must also mix-in a Capacitor and a Filter when using this strategy (see below).

The BoundedCapacityStrategy requires additional logic to function. Specifically it requires a Capacitor and a Filter. Capacitors are used to determine the pressure that the pool is under and provide a (usually) raw reading of this information. Currently we provide for the use of either mailbox backlog or active futures count as a means of evaluating pool pressure. Each expresses itself as a simple number - a reading of the number of actors either with mailbox sizes over a certain threshold or blocking a thread waiting on a future to complete or expire.

Filtering¶

A Filter is a trait that modifies the raw pressure reading returned from a Capacitor such that it drives the adjustment of the pool capacity to a desired end. More simply, if we just used the pressure reading alone, we might only ever increase the size of the pool (to respond to overload) or we might only have a single mechanism for reducing the pool size when/if it became necessary. This behavior is fully under your control through the use of Filters. Let’s take a look at some code to see how this works:

trait BoundedCapacitor
{
      def lowerBound:Int
      def upperBound:Int

      def capacity(delegates:Seq[ActorRef]):Int =
      {
              val current = delegates length
              var delta = _eval(delegates)
              val proposed = current + delta

              if (proposed < lowerBound) delta += (lowerBound - proposed)
              else if (proposed > upperBound) delta -= (proposed - upperBound)

              delta
      }

      protected def _eval(delegates:Seq[ActorRef]):Int
}

trait CapacityStrategy
{
      import ActorPool._

      def pressure(delegates:Seq[ActorRef]):Int
      def filter(pressure:Int, capacity:Int):Int

      protected def _eval(delegates:Seq[ActorRef]):Int = filter(pressure(delegates), delegates.size)
}

Here we see how the filter function will have the chance to modify the pressure reading to influence the capacity change. You are free to implement filter() however you like. We provide a Filter trait that evaluates both a rampup and a backoff subfilter to determine how to use the pressure reading to alter the pool capacity. There are several subfilters available to use, though again you may create whatever makes the most sense for you pool:

• BasicRampup - When pressure exceeds current capacity, increase the number of actors in the pool by some factor (rampupRate) of the current pool size.
• BasicBackoff - When the pressure ratio falls under some predefined amount (backoffThreshold), decrease the number of actors in the pool by some factor of the current pool size.
• RunningMeanBackoff - This filter tracks the average pressure-to-capacity over the lifetime of the pool (or since the last time the filter was reset) and will begin to reduce capacity once this mean falls below some predefined amount. The number of actors that will be stopped is determined by some factor of the difference between the current capacity and pressure. The idea behind this filter is to reduce the likelihood of “thrashing” (removing then immediately creating...) pool actors by delaying the backoff until some quiescent stage of the pool. Put another way, use this subfilter to allow quick rampup to handle load and more subtle backoff as that decreases over time.

Examples¶

class TestPool extends Actor with DefaultActorPool
                               with BoundedCapacityStrategy
                               with ActiveFuturesPressureCapacitor
                               with SmallestMailboxSelector
                               with BasicNoBackoffFilter
{
   def receive = _route
   def lowerBound = 2
   def upperBound = 4
   def rampupRate = 0.1
   def partialFill = true
   def selectionCount = 1
   def instance = actorOf(new Actor {def receive = {case n:Int =>
                                                   Thread.sleep(n)
                                                   counter.incrementAndGet
                                                   latch.countDown()}})
}

class TestPool extends Actor with DefaultActorPool
                               with BoundedCapacityStrategy
                               with MailboxPressureCapacitor
                               with SmallestMailboxSelector
                               with Filter
                                 with RunningMeanBackoff
                                 with BasicRampup
{
  def receive = _route
  def lowerBound = 1
  def upperBound = 5
  def pressureThreshold = 1
  def partialFill = true
  def selectionCount = 1
  def rampupRate = 0.1
  def backoffRate = 0.50
  def backoffThreshold = 0.50
  def instance = actorOf(new Actor {def receive = {case n:Int =>
                                                  Thread.sleep(n)
                                                  latch.countDown()}})
}

Taken from the unit test spec.