Scribd
Upload
16 views

Lecture Notes 3 - Scheduling

Uploaded by

chinnianu668
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
16 views

Lecture Notes 3 - Scheduling

Uploaded by

chinnianu668
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 13

III.

SCHEDULING

(a) General Concepts

- servicing of queues in network requires scheduling


discipline to determine order that packets serviced

eg. queues on network links, queues for servers

 scheduler responsible for managing delays of


packets and discarding of packets when queue full

- most obvious approach: first-come, first-serve (FCFS)

- in general, packets associated with different connections


or classes of service can have different priorities,
delays, and loss requirements

- conceptually:

SCHED 1
- 2 general service categories:

(1) guaranteed service


 network resources reserved to achieve
performance bounds
 "quality of service" (QoS)

(2) best effort


 no reservation of network resources required
 fairness in resource allocation critical

Typical QoS Parameters

Bandwidth: - minimum specified

Delay: - could specify worst case upper bound


or average case upper bound

Loss: - upper bound on fraction of packets


lost

Delay Jitter: - upper bound on difference between


largest and smallest delays

SCHED 2
- delay jitter critical in audio and video playback
applications where constant information required

- jitter can be removed at receiver by using "elasticity"


buffer but larger jitter  large buffers

 desirable to minimize jitter in network

(b) Priority Scheduling

- obvious way to give connections levels of service

- approach: logically put packets into a queue based on


priority level

- discipline: serve packet from queue n unless empty, then


serve packet from queue n-1 unless empty, then serve
packet from queue n-2, etc.

- problem: low priority queues can be starved if higher


priorities take all server's time

SCHED 3
- very easy to implement in both hardware and software but
cannot guarantee quality of service

(c) Work-Conserving vs. Non-Work-Conserving


Scheduling

- scheduler is "work-conserving" if only idle when queues


empty

 decreasing delay for queue i by giving more service


to queue i will increase delay for other queues

 scheduling discipline can only trade-off delay


between queues

SCHED 4
- scheduler is "non-work-conserving" if server may be idle
when queues not empty

 NWC scheduling can be utilized to minimize delay


jitter and to reduce buffers required for small
packet loss by smoothing out traffic

(d) Max-Min Fair Sharing

- one technique to share resource that satisfies users with


small demands and distributes remaining resource
evenly to large demand users

- assume resource with capacity C and n users with


demands d1, d2, ... , dn (in ascending order)

- procedure:

 resources allocated in order of increasing demand


so user gets a share of no more than 1/k of
remaining capacity of resource if k users left

 no user gets a share greater than its demand

 users with unsatisfied demand get an equal share

eg. - user 1 gets up to C/n of resource

- if d1 > C/n, then user 1 gets C/n of resource and user


2 gets up to C/n of resource

SCHED 5
- if d1 < C/n, then user 1 gets d1 of resource and user 2
gets up to (C-d1)/(n-1) of resource

etc.

- no user gets more than demanded or, if demand not met,


no less than any source with higher demand

- "max-min fair"  maximizes minimum share of a user


whose demand not satisfied

- can give users weights w1, w2, ... , wn to reflect relative


share (i.e., reflects priority of resource use)

- now limit on allocated share is in proportion to weight

Example:

SCHED 6
(e) Scheduling Best-Effort Connections

- fairness critical, i.e., max-min fair sharing desirable

Generalized Processor Sharing (GPS)

- idealized mechanism that achieves max-min fair sharing

SCHED 7
- packets go into a logical queue associated with their
connection

- each queue visited in round robin fashion and served for a


very small increment of time t if queue not empty

- as t  0, achieves max-min fair share

- can weight queue i with wi and serve for time wi  t in


each rotation

- GPS unimplementable!
 How can we serve a portion of a packet in a time
t  0?

Example:

SCHED 8
Weighted Round Robin

- similar to GPS except t  0 but serve an entire packet in


a rotation

- connections (i.e., queues) can have weights

- emulates GPS well for small, fixed size packets


(eg. ATM)
 over long enough periods of time very fair

What if packet size not constant?


 normalize connection weights by dividing by mean
packet size
 difficult to know mean packet size!

Weighted Fair Queuing

- basic concept:
 compute "times" to finish serving packets with
GPS server and then serve packets in order of
finishing "times" (actually "finishing numbers")

- let finish number = number used to reflect time at which


packet finished (but not really finish time)

- consider a GPS server where t = 1 bit time

- one round = serving one bit from all "active" connections


(i.e., one round not constant time)

SCHED 9
- an active connection has largest finish number of packet
waiting in queue or currently being served
> current round number

 time for one round = # active connections  t

FN calculation:

- packet arriving at inactive connection (i.e., empty queue):


FN = current round number + # bits in packet

- packet arriving at active connection (i.e., non-empty


queue):

FN = largest FN of packet in queue + # bits in packets

- FN does not depend on future arrivals, i.e., once


computed does not change

- round number does not increase at a constant rate with


respect to time but at a rate inversely proportional to
# active connections

SCHED 10
RN = t  link rate / # active connections + constant
(bits)

 RN vs. time = piece-wise linear graph

- when RN = FN indicates packet "done" according to


simulated GPS servicing

- in real scheduling, packets are served in order of FN

- do not need to view RN as # rounds for bit-by-bit round


robin server

 can view as real-valued variable proportional to


# active connections, i.e., RN just an abstraction

- complexity not in determining FN given RN but in


determining RN

See WFQ example.

SCHED 11
- buffer drop policy: when packet comes into queue, if
necessary, packets with largest finishing numbers can
be dropped to make room for packets

- for "weighted" fair queuing:


 assume wi = weight of i-th connection

- WFQ implemented in routers and ATM switches

(f) Scheduling Guaranteed-Service Connections

- can use WFQ to provide bandwidth and worst-case delay


bounds

eg. bandwidth allocated to connection i on a link can be


determined by

wi
 link rate
 j wj

SCHED 12
- delay jitter: WFQ not directly useful but one non-work-
conserving approach:

- can hold packets in regulator queue for connection i so


that packets do not arrive at scheduler any faster than a
specified rate

- evens out delays during packet bursts so that total end-to-


end delay more consistent

SCHED 13

You might also like