Outline

1
Outline

• Distributed scheduling
• Motivations
• Design issues
• Distributed scheduling algorithms

2
Motivations

• In a locally distributed system, there is a good
  possibility that several computers are heavily
  loaded while others are idle or lightly loaded
• If we can move jobs around (in other words,
  distribute the load more evenly), the overall
  performance of the system can be maximized

3
Motivations cont.
4
Motivations cont.
5
Distributed Scheduling

• A distributed scheduler is a resource management
  component of a distributed operating system that
  focuses on judiciously and transparently
  redistributing the load of the system among the
  computers to maximize the overall performance

6
Issues in Load Distribution

• Load estimation
• Queue lengths
• CPU utilization
• Load distributing algorithms
• Static
• Dynamic
• Adaptive

7
Issues in Load Distribution cont.

• Load balancing vs. load sharing
• Load sharing tries to reduce the likelihood of an
  unshared state (where one computer is idle while
  at the same time others are overloaded) by
  transferring tasks
• Load balancing algorithms attempt to equalize
  loads at all computers

8
Issues in Load Distribution cont.

• Preemptive vs. non-preemptive transfers
• Preemptive task transfers involve the transfer of
  tasks that are partially executed
• This transfer is in general expensive as it needs
  to transfer the entire task state consisting of a
  virtual memory image, a process control block,
  unread I/O buffers and messages, file pointers,
  times that have been set, and so on
• Non-preemptive transfers involve the transfer of
  tasks that have not started yet
• Environment transfer

9
Components of a Load Distributing Algorithm

• Four components
• Transfer policy
• Determines when a node needs to send tasks to
  other nodes or can receive tasks from other nodes
• Selection policy
• Determines which task(s) to transfer
• Location policy
• Find suitable nodes for load sharing

10
Components of a Load Distributing Algorithm
cont.

• Four components continued
• Information policy
• Demand-driven
• Periodic
• State-change driven

11
Stability

• The queuing-theoretic perspective
• The CPU queues grow without bound if arrival rate
  is greater than the rate at which the system can
  perform work
• A load distributing algorithm is effective under
  a given set of conditions if it improves the
  performance relative to that of a system not
  using load distribution
• Algorithmic stability
• An algorithm is unstable if it can perform
  fruitless actions indefinitely with finite
  probability
• Processor thrashing

12
Sender-Initiated Algorithms

• In sender-initiated algorithms, an overloaded
  node initiates the load distribution
• Transfer policy
• Selection policy
• Location policy
• Random
• Threshold
• Shortest
• Information policy

13
Sender-Initiated Algorithms cont.
14
Sender-Initiated Algorithms cont.

• Performance analysis
• Instability at high system loads
• When system loads are high, the sender-initiated
  algorithms can cause the systems to be unstable
• At high system loads, no node is likely to be
  lightly loaded and the probability that a sender
  will find a receiver is very low
• However, the polling activity increases as the
  rate at which work arrives increases
• Performance at low system loads

15
Receiver-Initiated Algorithms

• In receiver-initiated algorithms, an under loaded
  node initiates the load distribution
• Transfer policy
• Selection policy
• Location policy
• Information policy

16
Receiver-Initiated Algorithms cont.
17
Receiver-Initiated Algorithms cont.

• Performance analysis
• At high system loads, the probability of finding
  a sender is high and thus a sender can find a
  receiver in a few polls in general
• At low system loads, there are few senders but
  more receiver-initiated polls these polls do not
  cause system instability as spare CPU cycles are
  available
• A drawback
• Most transfers will be preemptive and thus
  expensive

18
Empirical Comparison of Sender-Initiated and
Receiver-Initiated Algorithms
19
Symmetrically Initiated Algorithms

• In symmetrically initiated algorithms, both
  senders and receivers search for receivers and
  senders respectively for task transfers
• The above average algorithm
• Transfer policy
• Location policy
• Sender-initiated component
• Receiver-initiated component
• Selection policy
• Information policy

20
Symmetrically Initiated Algorithms cont.

• Sender-initiated component
• A sender broadcasts a TooHigh message, sets a
  TooHigh timeout alarm, and listens for an Accept
• A receiver that receives a TooHigh message
  cancels its TooLow timeout, sends an Accept
  message to the sender, and increases its load
  value
• On receiving an Accept message, if the site is
  still a sender, choose the best task to transfer
  and transfer it
• If no Accept has been received before the
  timeout, it broadcasts a ChangeAverage message to
  increase the average load estimates at the other
  nodes

21
Symmetrically Initiated Algorithms cont.

• Receiver-initiated component
• It broadcasts a TooLow message, set a TooLow
  timeout alarm, and starts listening for a TooHigh
  message
• If TooHigh message is received, it cancels its
  TooLow timeout, sends an Accept message to the
  sender, and increases its load value
• If no TooHigh message is received before the
  timeout, the receiver broadcasts a ChangeAverage
  message to decrease the average at other nodes

22
Symmetrically Initiated Algorithms cont.

• Performance analysis
• Instability at high system loads
• Due to the sender-initiated components

23
Comparison
24
Adaptive Algorithms

• A stable symmetrically initiated algorithm
• Each node keeps of a senders list, a receivers
  list, and an OK list
• By classifying the nodes in the system as
  Sender/overloaded, Receiver/underloaded, or OK
  using the information gathered through polling

25
A Stable Symmetrically Initiated Algorithm cont.

• Sender-initiated component
• The sender polls the node at the head of the
  receiver
• The polled node moves the sender to the head of
  its sender list and sends a message indicating it
  is a receiver, sender, or OK node
• The sender updates the polled node based on the
  reply
• If the polled node is a receiver, it transfers a
  task
• The polling process stops if its receivers list
  becomes empty, or the number of polls reaches a
  PollLimit

26
A Stable Symmetrically Initiated Algorithm cont.

• Receiver-initiated component
• The nodes polled in the following order
• Head to tail of its senders list
• Tail to head in the OK list
• Tail to head in the receivers list

27
A Stable Sender-Initiated Algorithm

• This algorithm uses the sender-initiated
  algorithm of the stable symmetrically initiated
  algorithm
• Each node is augmented by an array called the
  statevector
• It keeps track of its status at all the other
  nodes in the system
• It is updated based on the information at the
  polling stage
• The receiver-initiated component is replaced by
  the following protocol
• When a node becomes a receiver, it informs all
  the nodes that are misinformed

28
Comparison
29
Performance Under Heterogeneous Workloads
30
Selecting a Suitable Load Sharing Algorithm

• The best algorithm depends on the system under
  consideration
• For example, if the system never attains high
  loads, sender-initiated algorithms will give an
  improved algortihm
• Stable scheduling algorithms should be used for
  systems that can reach high loads
• For systems with heterogeneous work loads,
  adaptive stable algorithms are preferable

31
Other Requirements of Load Distributing

• Scalability
• The algorithm should work well in large
  distributed systems
• Location transparency
• Determinism
• Preemption
• Heterogeneity

32
Case Studies

• The V-System
• The Sprite system
• Condor system
• The Stealth distributed scheduler