Partial-order transport service for multimedia and other applications

TLDR: Two metrics based on e/sub i/(P), the number of linear extensions of partial-order P in the presence of i lost objects, are proposed as complexity measures of different combinations of partial order and reliability.

Abstract: Investigates a partial-order connection (POC) service/protocol. Unlike classic transport services that deliver objects either in the exact order transmitted or according to no particular order, POC provides a partial-order service, i.e. a service that requires some, but not all objects to be received in the order transmitted. Two versions of POC are proposed: reliable, which requires that all transmitted objects are eventually delivered, and unreliable, which permits the service to lose a subset of the objects. In the unreliable version, objects are more finely categorized into one of three reliability classes depending on their temporal value. Two metrics based on e/sub i/(P), the number of linear extensions of partial-order P in the presence of i lost objects, are proposed as complexity measures of different combinations of partial order and reliability. Formulae for calculating e/sub i/(P) are derived when P is series-parallel. A formal specification of a POC protocol, written in Estelle, is presented and discussed. This specification was designed and validated using formal description tools and provides a basis for future implementations. >

Figures

Figure 9: Experiment 4

Figure 7: Experiment 1: Random POs

Figure 4: Parallel-Streams POs

Figure 3: Chain-Singleton POs

Figure 6: Experiment 1 : Parallel-Streams POs

Figure 5: Experiment 1: Chain-Singleton POs

Full text

OPTIMIZING PARTIALLY ORDERED TRANSPORT
SERVICES FOR MULTIMEDIA APPLICATIONS

RAHMI MARASLI PAUL D. AMER PHILLIP T. CONRAD
Computer and Information Science Department
University of Delaware, Newark, DE 19716 USA
Email:
f
marasli,amer,pconrad
g
@cis.udel.edu
References [1, 5] introduce a transp ort proto col that oers partially ordered service
for multimedia applications. This paper investigates howmuch the selection of a linear
extension aects system performance in a partially ordered service. We rst showhowto
identify better linear extensions of a partial order, and then determine the performance
gains by using such linear extensions at the time of transmission. To quantify linear
extensions of a partial order, we propose a new metric (
pBuf
-metric) that is derived
from buering probabilities. Since
pBuf
-metric is complex to calculate, a simplied
version called

-metric is also investigated. An OPNET simulation shows that for
certain partial orders, a linear extension optimized according to these metrics provides
some delay, and signicant buer utilization improvements over a non-optimal linear
extension. Thus, prudent transmission order selection in a partially ordered service
does improve system p erformance. Results also show that, in general,

-metric is as
eectiveas
pBuf
-metric in identifying better linear extensions of a partial order.
1 Intro duction
Computer networks traditionally oer either ordered (e.g., TCP) or unordered (e.g.,
UDP) transp ort service. Some applications suchasmultimedia do not need an or-
dered service since they can tolerate some reordering in the delivery of the ob jects.
The degree of reordering should be within the sp ecic limits of the applications;
otherwise problems result at the application layer such as increased complexity, in-
creased buering, and loss of synchronization. For such applications, neither ordered
nor unordered service is a perfect t. Ordered service insists on delivering all data
in sequence even if it results in higher delays and buer utilization. Unordered ser-
vice, on the other hand, minimizes delay and buer utilization, but provides no order
guarantees. If an application with some order constraints uses an unordered trans-
port service, the application programmer is burdened with the task of implementing
mechanisms for ob ject ordering.
To achieve b etter tradeos between order and other quality-of-service (QoS)
parameters, and to satisfy the minimal order requirements of applications, partially

This work supp orted, in part, by the National Science Foundation (NCR-9314056), the US
Army Communication Electronics Command (CECOM), Ft. Monmouth, the US Army Research
Oce (DAAH04-94-G-0093, DAAL03-92-G-0070), and the US Department of the Army, Army Re-
search Lab oratory under Co operative AgreementDAAL01-96-2-0002 Federated Lab oratory ATIR-
P Consortium.

ordered transp ort service has b een prop osed [1, 3 , 5]. Partially ordered service lls
the gap between ordered and unordered service by allowing multimedia applications
to specify the delivery order of ob jects in the form of a partial order. Since partially
ordered service do es not insist on delivering all ob jects in sequence, it can provide
lower delays and buer utilization than ordered service, while, at the same time,
guaranteeing a multimedia application's partial order requirements.
Analytic results from [8] show that, under particular network conditions, a par-
tially ordered service provides delay, buer utilization, and buering time improve-
ments over an ordered service. In a partially ordered service, for a given partial
order, any of p otentially manyvalid orderings of the ob jects (i.e., any linear exten-
sion) is permitted as a transmission order. Results also show that the actual choice of
which linear extension to use at transmission time aects the degree of improvement
of dierent p erformance statistics (e.g., delay, buer utilization, buering probabil-
ities). In general, p erformance improves as the distance b etween dep endent ob jects
in the sender's transmission order increases. These results suggest that for a given
set of network conditions, there exists an
optimal
(or a set of equally optimal) linear
extension(s) that will result in the lowest buer utilization, or lowest delay, etc.
This pap er investigates (1) ways of identifying optimal or near-optimal linear
extensions, and (2) the eects on overall system performance of using such linear
extensions as a transmission order. Unfortunately, with the current state-of-the-art,
determining the optimal linear extension requires performing a simulation experiment
for every possible linear extension of the partial order b eing considered. This is
clearly impractical. Thus, we rst investigate how to identify near-optimal linear
extensions in a practical way. For this, we prop ose a new metric (
pBuf
-metric) as a
means of quantifying a linear extension's go odness. This metric is based on buering
probabilities and is derived from analytic results [8]. Since
pBuf
-metric has a complex
expression, we also prop ose another metric (

-metric) which is a simplied version
of
pBuf
-metric and easier to compute.
By optimizing linear extensions according to these two metrics, weinvestigate
byway of simulation the signicance of the performance improvements obtained by
using near-optimal linear extensions over sub optimal ones. These results are helpful
to the users of partially ordered services in deciding whether it is even worthwhile
to seek near-optimal linear extensions as transmission order. We also compare the
performance gains that occur from using a linear extension optimized by
pBuf
-metric
with that of

-metric. Such results are helpful in deciding which metric to use in
practice in nding go od linear extensions.
The paper is organized as follows: Section 2 motivates a partially ordered service
through example applications.
pBuf
-metric and

-metric are introduced in Section 3,
and the simulation study is presented in Section 4. Section 5 summarizes the main
results and discusses future work.
2 Why Use a Partially Ordered Service?
References [1, 5] introduce the development and motivation for a partially ordered
protocol/service including several examples. For completeness, a summary of these
2

Figure 1: Screen Refresh
Consider an application that must do a \screen refresh" on a workstation screen/display
containing multiple windows (see Figure 1). In refreshing the screen from a remote
source, ob jects (icons, still or video images) that overlap one another should b e re-
freshed from b ottom to top for optimal redisplay eciency. Ob jects that do not
overlap maybe refreshed in any order. Therefore, the wayin which the windows
overlap induces a partial order.
Consider the four cases in Figure 1. A sender wishes to refresh a remote display
that contains four active windows (ob jects) named
f
1234
g
. Assume that the win-
dows are transmitted in numerical order and that the receiving application refreshes
windows as soon as the transp ort layer delivers them. If the windows are congured
as seen in Figure 1.A, an ordered service (sometimes referred as a FIFOchannel) is
required. In this case, only one ordering is p ermitted at the destination. If window2
is received before window 1, the transp ort layer must buer window 2 and deliver it
only after window 1 arrives and is delivered.
At the other extreme, if the windows are congured as in Figure 1.D, an un-
ordered service would suce. Here any of 4! delivery orderings would satisfy the
application since the four windows can be refreshed in any order. Each of these or-
derings represents a linear extension (
LE
) of the partial order (
PO
). As notation,
four ordered ob jects are written 1

2

3

4, and unordered ob jects are written
3

using a parallel operator: 1
jj
2
jj
3
jj
4(
x
jj
y
means there is no dependency relation b e-
tween ob jects
x
and
y
). Figures 1.B and 1.C demonstrate two (of many) window
congurations that call for a partial order delivery service. In these cases, two and
six linear extensions, resp ectively, are permitted at the destination.
2.2 Partially Ordered Service for Remote Do cument Retrieval
Reference [4] describ es a prototyp e system for the retrieval and displayofmultimedia
documents from a remote server using Partial Order Connection version 2 (POCv2),
a partially ordered and partially reliable
a
transport proto col providing coarse-grained
synchronization support. In this system, multimedia documents with temp oral char-
acteristics are described using a Prototype Multimedia Sp ecication Language (PM-
SL). This language gives the author the ability to express the synchronization, order,
and reliability requirements of the ob jects that make up a temporal multimedia do c-
ument. The application serving these do cuments can extract the order, reliability
and synchronization requirements from such a specication and communicate them
to the transp ort layer, which then provides the necessary supp ort.
This simplies application development, since the do cument display client need
not contain complex mechanisms for ob ject synchronization and reordering. It also
allows for graceful degradation, since the do cument can be presented \p erfectly"
when network conditions allow, and in a less than perfect but nevertheless acceptable
manner when network service degrades. Finally, the use of partial order and partial
reliability rather than ordered/reliable or unordered/unreliable service allows better
QoS tradeos b etween order/reliability and other parameters such as delay, buer
utilization and throughput.
The software that parses and enco des a PMSL document for transmission chooses
a linear extension of the partial order as the transmission order. In choosing this
linear extension various factors must b e considered, including the duration of the
individual multimedia objects, their synchronization relationships, and the impact on
performance. Therefore, the development of techniques for determining the relative
performance of various linear extension alternatives is useful to the developmentof
such systems.
3 Do es the Choice of Linear Extension Matter?
In a partially ordered service, the transp ort sender is permitted to transmit ob jects
in any order that does not violate the partial order [3 ]. Results from [8] show that
the choice of which linear extension (
LE
) is used by the sender can have signicant
impact on exp ected p erformance. In general, as the distance between dep endent
ob jects increases, expected p erformance improves. Intuitively, this result can be
a
Partial reliability refers to the notion that individual ob jects mayhave dierent QoS require-
ments with respect to loss; some may require guaranteed no-loss transport service, while for oth-
ers, best-eort transport service may suce. Partially reliable transport service provides a middle
ground between these two in which the loss tolerance of each ob ject can be specied individually.
References [1, 3, 4] consider partial order and partial reliability in juxtaposition, while [8] and this
paper fo cus solely on partial order.
4

explained as follows: given ob jects
a
and
b
such that
a

b
in partial order
PO
,by
increasing the separation of
a
from
b
, the exp ected time that
b
will b e buered due
to the network's loss of
a
will decrease. This is illustrated in Figure 2 which depicts
a scenario for ve ob jects where the rst transmission fails. In this example scenario,
the
LE
1
(i.e., \
abcde
") of
PO
=((
a

b
)
jj
(
c

d
))

e
buers ob jects longer than
the
LE
2
(i.e., \
acbde
") of
PO
. Notice that the
LE
2
's improvement is provided by
increasing the distance b etween ob jects
a
and
b
. For the simple scenario of Figure 2
(i.e., the scenario where only the rst transmission fails), it is easy to nd the b etter
LE
sof
PO
. On the other hand, when the p ossibility of losing any of the ob jects
is considered, identifying go od linear extensions or more imp ortantly, optimal linear
extensions, is more dicult.
Buffering time
for b
B.
LE_2
Deliver c
Buffer b
Deliver a, b and e
Sender
Receiver
Retransmission
a
e
c
b
d
a
Deliver d
Buffer e
A.
Buffering time
for b
Buffer b
Deliver c
Deliver a, b and e
Sender
Receiver
Retransmission
a
e
b
c
d
a
Deliver d
Buffer e
timeout
period
LE_1
Partial Order:
ab
d
e
c
Figure 2: Dierent linear extensions result in dierent buering times
How can we nd the optimal
LE
that maximizes system performance? One
possible way is simulating every linear extension, and choosing the one with the best
performance. Obviously, this is impractical b ecause of the time needed to simulate
every
LE
of
PO
. Even a small
PO
with just 10 ob jects can haveupto3
;
648
;
800
linear extensions.
To nd a goo d
LE
in a reasonable time, we prop ose two new metrics (
pBuf
-
metric and

-metric) that are designed to predict the expected performance of die-
rent linear extensions of
PO
s. Given two
LE
s of a
PO
, the b etter one can be
determined by comparing their
pBuf
-metric (or

-metric) values, instead of their
simulation result values. Thus, these predictors will be extremely useful by the
transport sender in selecting a b etter transmission order.
Both of these metrics are based on buering probabilities and derived from
analytic results [8]. Through these metrics, we hop e to identify the
LE
of a
PO
that
results in improved, if not optimal, system performance. These metrics consider the
eects of dierent system parameters (e.g., loss rate, buer sizes) while quantifying
5