Using System Dynamics (SD) simulation to explore and derive
self-adaptive techniques for information systems
Derek J. Edwards, Alan D. Pengelly
British Telecommunications ple
BT Labs,
Martlesham Heath,
Ipswich,
Suffolk,
IP5 7RE
United Kingdom
ABSTRACT
Rapidly changing customer needs require greater flexibility from the service provider to maintain
or improve quality of service. This is particularly true of the telecommunications industry which
is highly competitive and subject to rapid evolution.
To meet the desired flexibility, a company’s information systems must be able to adapt quickly
to the environment. As manual reconfiguration is becoming increasingly ineffective, a solution is
to endow the Information System with the ability to autonomously adapt to change.
Information Systems architecture is essentially driven by the dynamics of market trends and
company structure. Simulation techniques, such as SD, that mirror associated causes and effects
give essential insight into the critical interactions between the Information System and the
environment.
This paper discusses how SD has been used to simulate an adaptive information system and its
environment, and ways in which the models can be analysed to derive the principles and criteria
for adaption.
455
Introduction
The management and control of today’s network and operational support systems presents a
significant challenge to Telecommunications companies. These systems generally have very
complex architectures and a high degree of heterogeneity. They are typically geographically
distributed and incorporate large databases, such that managing and operating these systems
accounts for a large proportion of a company’s running costs. Many of the maintenance and
support activities involve manually based processes which, as well as being resource intensive,
also provide a source of inefficiency and poor quality of service (QoS). Automation of these
activities would provide the dual benefits of reduced costs and improved QoS. The question is
how?
Over recent years there has been a growth of interest in the area of adaptive and evolving systems
(Morecroft & Sterman, 1994). A number of leading academic institutions and companies are
active in this field. In addition, there is much interest within the Systems Dynamics (SD)
community as regards how SD can be used to model such systems.
The work described in this paper has used SD to model a distributed database system which can
adapt itself on the basis of usage and performance criteria so that it is always optimised with
respect to its environment. Results from this work will be used to derive novel adaption
algorithms” and techniques.
Context
The model comprises three main parts - the information system (fig. 2b), the customer/market
influence on the system (fig. 2a), and the effect of the system on the customer/market (upper
fig. 2c). These form a feedback loop which models the evolution of the system.
The Distributed Information System
The Information system under study can be considered as either being used directly by customers
(e.g. automated service information) or used within a business process by people who service
customers (e.g. help desk or fault reporting). In either case, variants of the model can be
implemented to suit the situation. The model represented here relates to the latter example which
can be used as the general case.
At the time of writing, the model is in the UaeteAl Server n
early stages of development and will 7
evolve as work continues. The model Comms link
shows a simplified Server node and Client
of a distributed database system where
different nodes would be substituted .
according to the experimentation and Fig] Users x
evaluation required. Let us assume that we ’
456
are interested in the information response time (performance) for a particular group of users
“Users A’ (fig. 1). These are connected via a communications link to ‘Server n’. Also using
Server n are Users x (via different links). Users A and Users x will influence the performance
seen by Users A as they are using the same server.
Within the SD model (fig. 2b), communications performance (return of data) is represented by
the upper path, and server performance (query response time) by the lower path. The separation
of the two paths in this way may be thought unusual. We might normally expect to model more
literally, i.e. data return to follow on from the query process. However, as the model normalises
different query types in order to determine server response times, it would be difficult to separate “
these out in order to determine the data return pattern for specific users. Instead, since each query
type is known at initiation, it is far easier to deal with the communication response separately.
This will lead to model sample time lag error, but over the intervals we are interested in (usually
several hours), response time samples (at most a few minutes) will average out the affect of these
lags. Normally, the overall response time would be the combined outgoing query processing time
plus the return data time, hence the model ‘sums’ these. This form of model allows the individual
effects of server and communication link to be easily seen, rather than a server ‘backlog’ being
caused by a slow communication link. It might also be considered a reasonable represeritation of
retumn data being cached, freeing the server. If required, feedback could be implemented between
the communication link and the server flow.
An adaption algorithm is run if system loading causes performance (‘Quality of Service’) to be
reduced beyond a defined threshold for a significant period, in order to determine a possible
improved alternative Client / Server connectivity and data distribution. Each suggested Client to
Server connection can then be investigated with the model, firstly to validate the proposed
configuration, and secondly to improve the algorithm by seeing where anomalies might occur. |
Hence initially (fig.1), Users A may be connected to Server 1, which also services Users C. The
adaption algorithm may suggest that a better configuration is to connect Users A instead to
Server 2, which will also service Users D, The different server performance (Server 2) and user
loading (Users D) parameters would then be substituted. The intention would be to adjust the
system to the new configuration as quickly as desired according to specified criteria - perhaps
ready for the same load the next day, or in an emergency within a few minutes.
Customer / Market Influence
The model allows us to shape probable influences on the use of the Information System by taking
into account known trends, causes and effects or even simulating less predictable scenarios. The
changes in usage may be, for example, simply an altered work-pattern with regard to a particular
group of users, fewer or more users, or a major reorganisation. Any of these may be short or long
term and will influence the demands on the system which may require a reconfiguration. Within
a large organisation, this is typically an upgrade or retune performed locally, rather than a
rebalance of the utilisation of the system as a whole across the company. Also, manual correction
is often only implemented when the situation has seriously deteriorated. Automated methods
457
allow for timely adjustment and possible fine-tuning maintaining a system as close to optimal
performance as possible.
The model attempts to represent these subtle and not so subtle changes and the likely time
periods and effects involved, As well as these being input directly to the Information System
model, the results can be used to provide usage pattern input scenarios to the adaption algorithm
for development and test purposes.
Effect of the System on the Customer / Market
There have been a number of papers and articles referring to the effects of competition on service
demand (similar principles to HPS, 1994). This model assumes that the primary goal is to
maintain a high perceived quality of service to the customer (compared with competitors) and
that this will maintain or generate new demand which in turn will influence the use of systems
that support that demand. If the demand cannot be met, quality of service will be perceived to
fall. Various factors can influence what we call quality. However, we are assuming that provision
of timely and accurate information, together with a sensible price are the key factors. If we can’t
find the correct information or can’t access or provide it quickly enough, then this will affect
customer perception. If the service is too expensive it will generally not matter how ‘good’ it is.
Other factors can be plugged-in as desired.
This aspect of the model is based on the assumption that if the response to user required data is
too slow, then we will not respond quickly enough to the customer need and hence customer
perception will be poor, causing a potential migration away from this service. The effect of this
(eventually) is to modify our work-practice (as discussed earlier) requiring a system adaption.
Ideally we need early sight of performance problems and rapid rectification in order to minimise
the ripple effect through to the customer. In this way, the system will tune itself to the day-to-day
variable demands, as well as to the more gradual changes, insulating the customer from the less-
desirable effects, as far as is possible.
Price is influenced by the cost of provision of the service. An attempt is made to reflect the
difference in the effect of timely and relatively cheap automated adaption methods compared
with a slower acting manual approach. Hence both the timeliness and cost of the service are
affected by the method used for adaption. Accuracy is affected by the fact that there is a tendency
to use incomplete information or to be unaware of the latest details if long delays are incurred in
trying to obtain the data. These influences are therefore used as the key quality factors used
within the model.
The longer it takes to improve the system, the more likely will customer perception be affected
and correspondingly the way-the company responds. If the delay is relatively long, the company
is more likely to be biased towards a reactive, ‘fire-fighting’ approach. This can lead to
unpredictable demands on the information system which may cause undesirable affects (perhaps
quality oscillations). On the other hand, if fine-adjustment improvements can be made in
relatively short timescales, then the company can take a more measured strategic approach to
458
change, leading to more predictable information system demands. The aim is to create a closed-
loop model that covers the above influences in adequate detail and is sufficiently rigorous to
validate the approach, methods and results of adaption.
Conclusions
Work to date has shown that the model can be useful in evaluating the criteria and response time
for adaption. Trigger conditions for adaption have in the past been based on the exceeding of
service-time thresholds, however experiments suggest that a more ‘damped’ view of conditions
may on occasions be more appropriate depending on the tolerance to fluctuations in system
performance. Also, the effect on the end-customer is likely to follow a predictable time-lag.
Using these modelling techniques helps with identifying the performance measures that are
required, enabling performance monitors to be located appropriately on the actual system for use
by the adaption process. Figures derived from these experiments can help to determine the
optimal scheduling for adaption and the speed at which this should be undertaken. Development
of the algorithm will be based on the SD model experiments and the results applied to the model
for validation and improvement.
* The Adaption Algorithm. By using existing or derived performance data, non-linear
calculations are carried out to determine the most effective Client - Server connectivity to
improve overall system and/or individual user-group perceived performance. This will generally
result in some users being connected to altemative servers with the need for corresponding data
movement or replication. The currently implemented algorithm uses analytical limited
combinatorial test calculations on possible connections and data allocation to arrive at a near
optimal configuration. If there is little latitude in the system, a direct improvement may not be
possible. The results are generally a compromise between calculation time and accuracy.
Methods to reduce calculation time and maintain sufficient accuracy are paramount. Other, more
esoteric techniques are also being investigated.
References
Elizabeth Oxborrow, (1989). Databases and Database Systems : 2" Edition. Chartwell-Bratt,
ISBN 0-86238-237-8
HPS (1994). Introduction to Systems Thinking and Ithink. High Performance Systems, Inc.,
Hanover, NH. p 103
Morecroft, J. and Sterman J. (1994) - Modeling for Learning Organisations.
Productivity Press.
459
revert
wo: Business Facility
Business Response
(people and process adjustment)
‘Customer Perception
Business Response
Reorganisation
Strategic
Loading Pattern
Fig. 2a. This section of the model represents the business response to the need for
change and activities which are likely to influence the scheduling and severity of
demands on the information system.
a ‘hlarmation System 4a]
Comms Queue Comme 4
asia
@ iT : (a ve
lal O ova sd
Data Size A Data Size B
Sorvor Queue
ny
user arrivals
{o Server 1
‘mean service Server 1
Query Delay
‘Addtonal Load (7) transaction Rate A
Moan Arrivals A
Ec tranesoton Fates
Maan Artivals B Single Client-Server Performance.
Arrivals 8 Two data fragments
Fig. 2b. This section of the model represents the communication and server
performance effects based on Client activity. Mean Arrivals A and B represent
differing queries and rates for differing data fragment sizes.
461
O Feedback!
User / Market / Customer Perception rad a
User Quality a8 User Quality ‘Customer Quality 83
Customer Q Fi
User Q Flow omer aFiow |X Biliomet Guid
‘Quality to Customer
Quali to Usor
Perception Rate
Price gtaleness. «Competitor Influence
User Tolerance
\
fais) Adaption Method & Result a a |
‘Adaption Method
‘Auto Adaption
Manual Configuration
Configuration
2]
Feedback?
Fig. 2c. This section of the model represents the effects of Information System
performance on quality as perceived by users and customers. Customers are also
shown influenced by external factors. Changes in quality influence the need for
adaption and its timeliness, which in turn affect system performance by
reconfiguration.
462
CC BY-NC-SA 4.0