A Virtual Time CSMA Protocol for Hard Real Time ... - CiteSeerX

A Virtual Time CSMA

Protocol for Hard Real Time Communication 1

Wei Zhao Krithi Ramamritham Department of Computer and Information University of Massachusetts Amherst, MA 01003

Science

The most common communication network used in dis-

Abstract

tributed hard real time systems is the multiple access network. In this type of network, stations transmit messages

We study a virtual time CSMA time communication

systems

deadlines.

protocol,

clocks:

In this

a real time clock

each

node

to be equal to the real clock.

message

when

maintains

time clock.

The virtual

occurs if at any time two or more messages are transmitted

When-

clock

the time on the virtual

on the channel.

the minimum-laxity-first

No message can be received correctly in

the event of a collision.

then

clock

Only one message can be success-

fully transmitted over the channel at any time. A collision

two

Based on how the collisions are handled, multiple access

A node transmits

equal to the latest time to send the message. implements

via a shared channel.

have explicit

to be idle, it resets its virtual

runs at a higher rate than the real clock. a waiting

for hard real

where messages

and a virtual

ever a node finds the channel clock

protocol

communication protocols can be broadly divided into three

is

categories:

This protocol

transmission

policy.

1. lrzjerence

avoiding

protocols:

This category includes

We compare the performance of our protocol with two

ALOHA [1] and various CSMA protocols [9]. These

baseline protocols both of which transmit messages accord-

protocols operate without taking past history of the

ing to the minimum-laxity-first policy. While both use per-

channel into account.

fect state information about the nodes and channel, the 2. Inference

first is an idealized protocol which obtains this informa-

seeking

protocols:

Various tree, window,

and stack protocols [2,4,5,7,12,27] and Urn protocols

tion without paying any cost and the second one pays a

[10,28] belong to this category. These protocols make

reasonable price for it. The simulation study shows that

inference on the collision history, and usually solve

in most cases, our protocol performs close to the first one

collisions by partitioning some parameter space of

and better than the second one.

messages. 3. Deterministic

1

Introduction

or Collision-free

Division

Multiple-Access

Bit-Map

Protocol

with Alternative This paper concerns communication

protocols

for dis-

protocols:

Protocols

The Time

(TDMA),

the

[11], the Broadcast Recognition Priorities

Protocol

[24,3,6],

the

Multi-1evel Multi-access Protocol [23], etc. are in this

tributed hard real-time systems. Tasks performed in such

category. They work in such a way that collision

systems are time constrained. Consequently, the messages

not occur at all.

do

transmitted in the network often also have explicit time constraints such as deadlines.

and evaluate a protocol which particularly time constraints control

into account,

of communications

While we are currently developing protocols belonging to

In this paper, we propose takes messages’

and hence is suitable in distributed

all three categories for hard real-time communication,

hard

this

for the

longing

real-time

systems.

we report

to the

first

on the

performance

of a protocol

in be-

category.

In our model, each meseage has a deadline which specifies the absolute time by which the message

1This material is based upon work supported in part by the National Science Foundation under grants DCR8318776, and DCR-8500332, and by the Office of Naval Research under grant 048-716/3-22-85.

must arrive

at its destination; otherwise it is useless. The major performance metric in such a network is the ratio of message loss, i.e., the fraction of messages that do not arrive at

120 CH2351-5/86/0000/0120$ 01.00 @ 1986 IEEE

paper

their

destinations

by their

A good protocol for

deadlines.

results of simulation studies comparing the performance of

hard real-time communication should minimize this ratio.

our protocol with that of two baseline protocols.

The design objectives of multiple access protocols and

5 concludes the paper and suggests future work.

Section

scheduling algorithms are quite similar: Both are for allocation of a serially-used resource to a set of processes [13]. It is known from the tkeory of hard real-time scheduling

2

The Model

that, in the static case, i.e., when all the task characteristics are known a priori, minimum-deadline-first minimum-laxity-first

and the

In a multiple access network, a set of nodes are connected

scheduling policies are optimal in the

to one communication

sense that they can schedule a set of tasks if there is some

channel.

At any given time, only

one message can be successfully transmitted over the chan-

policy which can do so [15]. In the dynamic case, these

nel. The maximum end-t o-end delay for a bit is r.

policies also offer better performance than others [8]. Due

assume that the time axis is slotted.

to this fact, in our protocol,

We

The length of a slot

we adopt a network-wide

is defined to be one time unit. Given that the maximum

transmission policy in which the message with the mini-

end-to-end delay is r, we let the length of a slot be equal

mum laxity is transmitted first.

to r. A node can start transmitting a message only at the beginning of a slot. The length of a message is a multiple

Kurose et al. suggest a window protocol for real time communicantion [12,13], implement ing the minimum-laxity-

of the length of a slot. The

first policy. They assume that the laxities of all the mes-

defined as

sages when they arrive are constant. tion, the minimum-laxity-first first-come-first-served also studied They

the problem

assume

transmitted

policy.

that

policy is ident ical to the

the length

on the channel

transmission

where M is the mean message length.

is independent

delay

is

Previous work

shows that CSMA protocols are applicable to environments

policy.

of the i-th message

end-to-end

Q = TIM

Under this assump-

Panwar et al. [20,21] have

of optimal

normalized

where a has a small value such as 0.01.

being

Each message, M, can be characterized as follows:

of what the i-th

message is. In other words, the lengths of messages varies with the order in which they are transmitted. that if the channel is a message policy

waiting

is the best.

are removed:

In our protocol,

We allow messages

clock

and

to have arbitrary

lengths to be determined

each node maintains

a virtual

finds the channel

laxit ies, at the

Latest time to Send the message, LSM, is equal to

with the order

time

clock.

two clocks: Whenever

to be idle, it resets its virtual

to the real clock.

higher

Deadline DM. This is the time by which message M must be received by its destination;

DM – LM;

they are transmitted.

In our protocol,

equal

needed to transmit message M;

all of above assumptions

time they arrive and hence to be invariant

time

Length LM which is the total number of time uni&

prove

to be sent, the minimum-laxity-first

and also allow the message

in which

They

is not allowed to remain idle when there

The virtual

rate than the real clock.

clock

Laxity at time t,LAM(t)

a real a node

at time t.Therefore,

clock to be

then runs at a

A node transmits

is the maximum amount of

time the transmission of message M can be delayed

LAM(t)

= DM – LM – t = LSM – t.

a wait-

clock is equal to

When it is clear from the context, we may omit ar-

the latest time to send the message

which

is equal to the

gument t as well as subscript M in the above expres-

message

length.

This protocol

ing message

when the time on the virtual

deadline

implements Our protocol

cols — virtual discussion The

minus the message

the minimum-laxity-first belongs

to the newest class of CSMA

time CSMA

[18,19,14].

of the general virtual remainder

transmission

Section

2 defines the model

Section

3 introduces

in Section

is organized

we adopt

our protocol,

From these definitions, it is clear that if we transmit tasks

proto-

according to their latest start times, it is equivalent to the

We will have a brief

time CSMA

of this paper

sions.

policy.

3.

as follows:

in this study.

Section

minimum-laxity-first

While

4 presents

the

121

transmission policy.

3

A Virtual Time CSMA col: VTCSMA-L

b. Every node drops any waiting message if the LS

Proto-

of the message is less than the current time t, i.e., such a message is lost. 4. A node sends message M when VLSM

Various Carrier Sense Multiple Access (CSMA) protocols

equals the

have been proposed and evaluated [9]. Most of them may

reading of the virtual clock. Note that because the

not be suitable for hard real-time communication because

virtual clock runs faster than the real clock, the mes-

the message transmission delay cannot be bounded. Molle

sage is sent at (real) time

and Kleinrock [18,19] suggested a new class of CSMA protocols called Virtual node

Time

has two clocks.

another

One clock

gives the virtual

ning when the channel is idle.

CSMA.

time.

In this protocol,

each

gives the real time,

and

The virtual

where ~ > 1. This time is earlier than the latest time

clock stops run-

to send message M, LSM.

is busy, and runs when the channel

5. When a collision occurs during the transmission of a

When it runs, the virtual clock runs at a higher rate

than the real one if it is behind

the real clock.

time is equal to the time on

the virtual

advantages

protocol

clock.

The major

are its fairness

in transmitting

the waiting

rate, and better

[14].

we propose

real-time

run along

a virtual

communication.

Because

of this,

time

CSMA

LS such that

transmission

we call this protocol

policy

b. draws a number R randomly from the interval (t’, LSM),

protocol

We let the virtual

the axis of messages’

wide minimum-laxity-first

behavior

ability Pi; or

mes-

lower collision

Below,

a. re-transmits this message immediately with prob-

of the VTCSMA

sages (since it is based on the FCFS policy), delay-throughput

the sender node message, say at (real) time t’,

A message

is sent only when its arrival

hard

t + (VLSM – t)/~

for

VLSM

as

E R,

clock

a network-

c. then, puts this message back in the queue of mes-

is achieved.

VTCSMA-L.

and modifies VLSM

sages waiting to be transmitted.

The

following is the outline of our protocol:

The protocol above ensures that the message with the minimum VLS is transmitted first. By the definitions of

1. Each node has two clocks, one clock maintaining the

VLS, LS and LA, we notice that this policy is equivalent

real time, and another maintaining the virtual time

to that of the minimum-laxity-first.

set in a manner discussed below.

The following example helps to understand how the

2. Each message M is associated with a parameter virtual latest time to start

transmission,

VLSM.

time on the virtual clock relates to the time on the real clock (eee Figure 1).

when

a message arrives, VLSM of message M is set to be

EXAMPLE:

LSM. If the channel is idle and

found idle, every node resets its virtual clock to be t~= tl.

‘LsNew

Assume that at real time tl the channel is

Then the virtual clock runs at a rate of q > 1. Say, two

Message < – the reading of the virtual clock,

messages Ml and &fz are waiting to be sent with VLSM, =

the new message is sent immediately; otherwise it

tjand VLSM2 = tjrespectively.

waits. VLSM may be modified when the transmission

When the virtual clock

Ml is sent out. This message is transmitted equals to t;,

of this message causes a collision. We will discuss the

over the channel from the real time t2 to ts.During the

details of this modification in Step 5.

transmission of Ml, the virtual clock does not run.

3. Each node eenses the channel. Suppose that after ei-

At real time t~, because the channel is idle again, the

ther a successful message transmission or a collision,

virtual clock is set to be t~ = t3.Then it starts running at

at real time t every node finds that the channel is

rate q. When

idle:

sent out.

The channel

a. Every node resets its virtual clock to be equal to

running.

Say, at real time

the real clock.

the virtual

clock

reaches

t:, message

is busy, and the virtual

clock

&fz is stops

t5,the transmission of message

M2 finishes. Then the virtual clock is set to the value of

The virtual clock starts to run

and starts running again. the real clock (so t~= t5),

at rate ~ > 1. ~ is a parameter of this protocol.

From this example, it is clear that the virtual clock

We will discuss seIection of q later. The virtual clock will stop when the channel is busy.

does not run continuoust~.

122

At real time t3,the virtual

realizable in practice. We use it purely to give us an upper bound on performance. The second baseline protocol is based on the binary countdown protocol proposed by Mok and Ward [16]. We modify it as follows for use in hard real-time communication. 1. Let MLi

.----.,,// ,, ,, “r; /’:

‘/

‘/

. . .

to be sent, it sends out the complement

t

code of its MLi

of the binary

one bit per slot in a binary cormt-

down manner [16,26]. Briefly, the binary countdown method functions as follows: as soon as a node sees that a high-order bit position that is O in its MI/i has been overwritten with a 1 (by some other node(s)),

,,

P

..-.

/:

senses the channel to be idle, if it has any message

!,

,,

/:

.

/

be the minimum laxity (> O) of the mes-

sages waiting to be sent on node i. When each node

this node gives up and does not send any more bits.

,,

.

,’

/,6

2. After a complete

,,

(complement

code of) MLi

has

been transmitted along the channel, each node knows what the minimum laxity among all the messages waiting to be sent is. Those nodes, which have messages with this minimum laxity, then send out their identification number in the same binary countdown manner.

Figure 1: Virtual Clock vs. Real Clock clock is set to equal the real clock.

3. The node, which has the message with the minimum

It is not necessary

for the virtual clock to cover the interval (tj, t!Jbecause if

laxity and has the minir%um identification number,

t:)this message there is any message with its VLS in (tj,

starts sending out the message.

has been lost by real time ts.On the other hand, (t&,t’J is twice covered by the virtual clock.

4. Each node repeats the above procedure when the

This is necessary

channel is idle again.

because when message iW2 is being transmitted, some new messages may arrive with its VLS in that region and at

Because the “countdown”

real time ts,there is still a chance to transmit it.

is done on Iaxities of messages,

we call this baseline protocol Binary Countdown

on Lady,

abbreviated as BC-L. Note that BC-L uses

4

4.1

Performance Simulation Base-line

Centralized

[log, (Maximum Laxity)l + [log, (Maximum Node Identification Number)l time slots to decide which message is to be transmitted

Protocols

during each cycle of transmission.

We will compare the performance of VTCSMA-L baseline protocols.

by

Evaluation

with two

The first baseline protocol is called

Minimum

Lazit~

message

transmitted

4.2

Simulation

Results

First, We simulate 3 cases for a network of 10 nodes. The first

abbreviated as CMLF. In this protocol, all the transmissions of messages are assumed to be scheduled by a central-

two cases are for stochastic messages and the last one is

ized controller. This controller obtains perfect knowledge

for periodic messages.

about the nodes and the channel without any communica-

ten in Simscript, and runs in an ULTRIX environment on

tion overheads. It schedules the transmissions of messages

MicroVAX-II.

such that the message with the minimum laxity is trans-

probability of immediate re-transmission, FI, takes a value

mit ted first.

of 0.5. We will discuss extension to this in the last section.

This algorithm is hence an ideal one, not

123

The simulation program is writ-

For simplicity, in all the simulations, the

K-1

~ ~

T

vl[sHfl-

—

~

M-1

—

w

~

c

Ollf

0.80

!,!0

0,50

0.50

0.40

0.40

willl-1

0,11 0,?0

0,10 Loo~ ~ I 1 I I I t I I I I 0.00.10,7o,~’0,! 0,s0.6oil 0,80,91.01.11,1 ~

4.2.1

0.00

I

I

I 0.l !,?

I

I

I

I

I

I

I

I

Case 1: a = O.Ol, q = 27.0

Case 2: a = O.l, rl = 4.0

Figure’2: Message Loss vs. System Load

Figure 3: Message Loss vs. System Load

Stochastic

I

I

o,~0,40,50.60,10,80,91.01.11,1 ~

Messages

2. When a is small, e.g., equal to 0.01, the performance In Cases 1 and 2, messages process.

The arrival rate at each node is the same.

-lengths

are exponentially

in Case randomly 2. Recall Thus,

of VTCSMA-L

arrive at each node as a Poisson

distributed

with

1 and 10 in Case 2. The laxities

a mean

of 100

of messages

3. When a is large, e.g., equal to 0.1, the performance

are

of VTCSMA-L

chosen from [0, 600] in Case 1 and [0, 60] in Case that the end-to-end

in Case

the network

the system

L=

with system

load (L) is defined

the performance of VTCSMA-L

In

load from as

4. In any case, VTCSMA-L

Selection of q is an important issue for the application of VTCSMA-L.

number of nodes in the network. For the simulation, N = and =

10

is definitely better than

BC-L.

AR*M*N

node, M is the mean of the message length, and N is the 1,

is closer to that of

CMLF than to BC-L.

where AR is the mean of the message arrival rates at each

10, M = 100 in Case

is between that of CMLF and BC-L.

When the load is not too high, say, lower than 0.7,

delay r is the unit of time.

1, a = 0.01, and in Case 2, a = 0.1’.2

each case, we simulate 0.1 to 1.2 where

is almost identical, or very close to

that of. CMLF.

Message

For each value of CY,we test the system

performance over different values of q with three different

in Case 2. AR is adjusted

loads.

to get different values for L,

Figures 4 and 5 show the simulation results. We

see that when the system load is light (= 0.1 ),

Figures 2 and 3 show the ratios of message loss (R) vs.

is small

(= 0.01).

the system loads (L) for Cases 1 and 2 respectively. From

message

loss is very close to the minimum.

these figures, we can make the following observations:

that the system chosen

1. Message loss in CMLF and BC-L tends to grow al-

forms

most linearly with increase in system load.

[18,19],

‘We note that a = 0.01 is considered to be the typical application environment for CSMA [19].

performance

is not sensitive

performance are different.

a

This indicates to the values

This observation

with that made by Molle for the original though

or when

wide range of ~ values,

for q under these situations.

our VT CSMA-L

124

For a relatively

metrics

con-

VTCSMA

of his VTCSMA

and

Q

e

o ml ~ 0,9

0

e

B

d

b

El

0

0

1o11o ~0,5

10.!0 ‘ 0.1

!,60 0.50

A

0.s0

0.40 O,M \

0.30

0,)0

0.20

0,10

0.10I

0,!0

I

1

t

I

I

I

I

I

I

?.0 3.! ‘iiu 10,0 10,030.0 50,0 100,0 “ case 1: c1=

U.!(I 1.0

Periodic

Case 2: a =0.1

0.01

Figure 5: MessageLoss vs. q

IYlessages

Case 3 is for periodic messages.

Protocols BC-L VTCSMA-L CMLF

In hard real time sys-

tems, messages are often periodic, i.e., they arrive and request to be transmitted pertidically. be input/output

sill lit v

2,0 io S,u 10,0 20.0 N

Figure 4: MessageLoss vs. q 4.2~2

10RO :0,9 10IO ~0,5 10RO ‘ 0,1

These messages can

b= 01b=5751b 0.00 0.11