You are on page 1of 4

Key Issues of Resource Management in Distributed

Multimedia Computer Systems


Zhang Zhanjun Yang Xueliang Han Chengde
Graduate School of University of Science and Technology of China, Beijing, I00039
Institute of Computing Technology, Chinese Academy of Science, Beijing, I00080
2% I irLl ch pc . I c t . ac . c I1
Abstract: The guarantee of quality of service (QoS) of
multimedia applications is a kernel issue in resource
management of distributed multimedia computer
systems. Some key issues are done research for
resource management in multimedia transmission in
this paper as follows. ( I ) the relation between QoS
and resource management, (2) the algorithm of
scheduling distributed multimedia tasks, (3) the
paradigm of session-based distributed multimedia
resource allocation.
Key Words: QoS, scheduling tasks, session, resource
management.
1. Introduction
The guarantee of quality of service (QoS) of
multimedia applications is a kernel issue i n resource
management of distributed multimedia computer
systems. For example, a video stream requires QoS
including end-to-end delay 100ms. CPU, I/O and
buffers should be allocated appropriately for
guarantee of the delay in computer systems.
Bandwidth should be allocated appropriately to
guarantee the delay i n network systems. It is an
important issue how to allocate these resources. The
application's QoS can not be guaranteed if required
resources can not be allocated. The resources can not
be utilized effectively if excessive resources are
allocated so that the resources are wasted.
The relation between QoS and resource management
and the content of resource management are analyzed
in this paper. This paper focuses on the scheduling
tasks and resource allocation. This paper does
research for the algorithm of scheduling multimedia
tasks because the algorithm has impact on CPU
management. This paper does research for the
paradigm of resource allocation for CPU, I/O, buffers
and network bandwidth using the producer -
consumer model, and presents the paradigm of
session-based resource cquilibriuni allocation for
guarantee of steady flow.
2. Related Work
Most of algorithms of scheduling distributed
multimedia tasks are from real-time systems e.g. EDF
and RM. They are periodic scheduling algorithms.
The pinwheel scheduling is based on a Distance-
Constrained Task System (DCTS), unlike periodic
scheduling, enables scheduling no jitter so that it is
very suitable for the distributed multimedia systems.
Han presented a single-node pinwheel algorithm, Sr .
Hsueh extended Sr to the DSr algorithm so that
multiple nodes are processed. However, the definition
of DCTS of DSr algorithm is very simple so that it
can not be utilized in real multimedia systems.
Mehra discussed resource management in real-time
' This work was supported by the Chinese National Science Foundation 69983007.
0-7803-6394-9/00/$10.00O2000 IEEE.
systems. Huang presented resource management
model for continuous multimedia database
applications in uni-processor system, and studied a
paradigm of allocation of threads, I/O processes and
buffers. Nahrstedt presented the content and key
issues of resource management for network
multimedia applications.
3. QoS and Resource Management
CPU, memory (buffers), I/O (disk I/O and multimedia
device UO), bus bandwidth and network bandwidth
are shared resources in distributed multimedia
computer systems. The bus bandwidth is dynamically
allocated by operating system according to processes
or threads executing so that it can not be managed by
application process. T~L I S, it is not studied in this
paper. Some appropriate resources should be allocated
for guarantee of application's QoS. The application's
QoS can not be guaranteed if required resources can
not be allocated. For example, CPU, memory, I/O and
network bandwidth should be allocated for guarantee
of end-to-end delay in QoS in computers along end-
to-end path. Thus, resource management is relative to
the QoS and the guarantee of application's QoS is
premise of resource management.
Sometimes, The QoS of a few applications can not be
met as exhausted resources. For example, a video
stream requires a 640K byte buffer. But there is only
540K byte available memory in system. Thus, The
utilization of resources should be improved
effectively for guarantee of QoS of the more
applications'. When system allocates resources for
each application, excessive resources can not be
allocated
4. Multimedia Task Scheduling
4.1 Scheduling Algorithm in Real-Time System
The most of scheduling algorithm are periodic in real-
time system. The time constraints of the periodic task
Tare characterized by the following parameters T(s, e,
d, d r ) .
s: Starting point
e: Processing time of T
d: Deadline of T
p : Period of T
I-: Rate of T (i-=l/p)
where O$e<'d$p. The starting point s is the first
time when the periodic task requires processing.
Subsequently, it requires processing in every period
with a processing time of e. At s+(k-I)p, the task T is
ready for /-processing. The processing of T in period
k must be finished at s+(l;-l)p+d. For continuous
media tasks, it is assumed that the deadline of the
period ( k- I) is the ready time of period k .
(1)
The Earliest Deadline First (EDF) algorithm is one of
the best-known algorithms for real-time processing.
At every new ready state, the scheduler selects the
task 7; with the earliest deadline dl among the tasks {7?
that are ready and not fully processed. The requested
resource is assigned to the selected task. At any
arrival of a new task q, EDF must be computed
immediately leading to a new order: the running task
T, is preempted and the new task T, is scheduled
according to its deadline (Il. The new task T, IS
processed immediately if its deadline a'/ is earlier than
that of the interrupted task T,. The processing of the
interrupted task is continued according to the EDF
algorithm later on. Given a task set T={T,, q, ... TI ) ,
EDF is guaranteed to find a feasible schedule for T if
Earliest Deadline First Algorithm (EDF)
(2) Rate Monotonic Algorithm (RM)
(a) The rate monotonic scheduling algorithm is an
optimal, static, priority-driven algorithm for
preemptive, periodic jobs. The order of tasks is
computed before tasks in {TI are scheduled.
Subsequently, each task is processed with the priority
calcplated according to its rate at the beginning. The
faSter rate i i a task is, the higher its priority is. If tasks
TI and T, are given by u,<r,, their priority are defined
Given a task set T={T1, T,, ... TIS, RM is guaranteed
to find a feasible schedule for T if
by P/'l<Pr/.
I1
Ce i i p i I n ( 2 " " -1).
,=I
I . 2 Distributed Multimedia Task Scheduling
I n some real-time applications, task must be executed
i n a distance-constrained manner, rather than just
periodically. That is, the temporal distance between
any two adjacent executions of the same task should
always be less than a certain amount of time. Such
real-time systems are called Distance-Constrained
Task System (DCTS) [9]. The pinwheel scheduling is
based on the DCTS and it is very suitable for the
distributed multimedia systems with jitterless.
Definition 1: Let X={Xi) I <i g n , be a set of
distribh'ed multirriedia tasks. Suppose the number of
tasks (transactions) I S n, and the number of nodes is nz.
For YX E X , XI ={ T , , } , ~ E {l . . . Rz } ~ let T,,be the
task of X, executing on node NI. TI, has the distance
constraint c,, and executing time e,/.
Definition 2: Given a set of scheduling multimedia
tasks x as DCTS,
density of X in node NI .
Theorem 1: Given a set of scheduling multimedia
tasks X as DCTS, i fX is met in any nodes N,
Z , ( x ) 5 P=Pr ' ,
then DMSr is guaranteed to find a feasible for X;
where 17, is number of tasks in N, .Moreover, the
bound Bis tight.
5. Session-based Resource Allocation
Definition 3: There are n multimedia streams in
system. The multimedia stream can be described as
the session .$-( Ai , 78 g' 8, i=1,2 ,... , n [ I O].
.Ii-the target rate of media data flow specified from
the end-user perspective. It is defined by data units
per second, where a data unit can be a video frame or
a group of audio samples consisting of a certain
number of bytes.
; Fa set of threads q,(j=1,2 ,... in, m1 2 / ) that
execute 5 !i is defined by rj/ (4, 6 $, where 4 is
its execution period, $ is its CPU time or period for
processing one media data unit, 3 is the number of
data units processed within each period, and F; . . XTis
the executing time of thread r,?
p a set of buffers k(j=/, 2 ,...,I,
such as a video frame and $is the number of such
data units allocated i n the buffer. Hence, a buffer size
113/) allocated for +'
is defined by y, #, where z, j is a media data unit
is pq
+- an I/O process that carries out the I/O requests of $'
+is defined by Sy g q $, where Tis the I/O request
period, q is the unit of media data on device (disk),
and Tis the number of allocated units to be accessed
within each period p
F a set of network processes 7/(7=/,2,. ..., h,
that execute network transmission of $ yi is defined
by y j ( q, 5, $, where 4 is period of network
processing, 5 is transmitting unit in network, ai s
number of media units transmitted per period (3.
hj&)
Theorem 2: There are n sessions &,5( AI, 78 g' $<
i =/ , 2, ...,a i n system and there are one I/O process, nz,
threads, and h, network processes in session S,. The
period of I/O processing r,, the period of threads p, / , /
Sj </77,,, and period of network processing q,/ , / S'j S'
h, are allocated by the number of media units x,,
processed as follow.
Xl (X +1 1 1 1 )
XI ( h +,111 +I )
q l k = x q i ( h +I ) 1 I IC <hi
xi ( k +111, )
X I ( , +I )
q l k =- x pi( j +I IC =hi, j =m i +1
j = ml
1
=;k
Theorem 3: Given a session ?with m, threads, hl
network processes, and i nl +hl buffers. The utilization
of CPU VI, buffers MI, l/O bandwidth Z,, and network
bandwidth J, are allocated by the nuniber of media
units x, , processed as follow.
/=I
Ll i l XXl O
=- =auil
Ii
6. Conclusion
The guarantee of quality of service (QoS) for
distributed multimedia streams greatly challenges
resource management in computer systems and
network systems. Resources in networks and
computers should be allocated according to the QoS.
There is close relation between resource management
and QoS. The essential of resource management is the
guarantees of QoS. Most of algorithms of scheduling
multimedia tasks are from real-time system such as
EDF, RM, which are periodic. This paper presents a
distance constraint-based sclicduliiig algorithm DMSr
which pinwheel bases are calculated step by step in all
nodes in parallel. A paradigm of session-based
resource allocation is presented for steady rate of
multimedia streams in this paper. When resources are
allocated, it can calculate the resources for steady
continuous multimedia streams by making
equilibrium equations among process or thread s
periods, I/O processors periods, buffer space and
network bandwidth. Fuithermore, we will study
distributed session scheduling with adaptive network
bandwidth availability, for example Internet.
References
[l ] Marco Di Natale, J ohn A Stankovic, Dynamic
End-to-end Guarantees in Distributed Real-Time
Systems, Proceeding of Real-Time System
Symposium, 1994
[2] Hiroyuki Kaneko, J ohn A. Stankovic, et al, A
Multimedia Server on the Spring Real-Time System,
IEEE Real-Time Technology and Applications
Symposium, MA, J une, 1996
Zhang Zhanjuii received his BS i n computer science
from HeBei University in 1988 and his Ph.D in
computer science from University of Science and
Technology of China in 1996. Currently, he is a post-
doctor researcher. His current research interests
include the multimedia transmission, distributed
computer systems and network protocols.

You might also like