The m anagement re-configuration mechanisms were presented in detail in Section 4.3.1. It is important to investigate the consequences o f the extra load introduced by a re-configuration and
the subsequent restoration procedure. Appendix A, Section A.2.5 gives the equation for the extra signalling load generated during the restoration period that has to be processed by the nodes located in the hierarchical path between reconnected child and old parent. Equation (A .42) for the extra signalhng generated after a management re-configuration is repeated below.
For tlie above equation, rrio is tlie number o f database entries at reconnected cliild node when tlie re-configuration occurs and ripath is the number o f nodes tlie restoration packets have to traverse between reconnected child and old parent, inclusive. Hence, when choosing the nodes to be recormected, preference should be given to the ones at a low er hierarchical level - i.e. smaller number o f database entries, and tlie new parent node should be selected according to its proximity
to the old parent node so that ripath is restricted.
The m echanism for database restoration was designed so that im mediate restoration packets are
only em itted w hen the reconnected child has the spare capacity to process them (see Section 4.3.1.2). In that w ay the immediate restoration rate is dynamically adapted to tlie current load on the reconnected child and consequently the load on that part o f the network. If the immediate restoration rate is turned o ff com pletely due to the current load on the reconnected child, then only restoration on demand is performed till the restoration procedure is completed. Therefore, even w hen the re-configured portion o f the network is overloaded, the restoration procedure is performed in a smooth and transparent way.
Equation (A .42) does not take into account concurrent re-configurations involving the same portion o f the network. As explained in Section 4.3.1, restoration packets from concurrent re configuration procedures might clash with one another leaving the corresponding data path corrupted. This consequently causes a recovery m echanism to be performed when the inconsistency is detected, increasing the signalling load. The interactions am ong concurrent re configuration procedures and their consequences in terms o f request throughput and end-to-end delay need to be investigated using the simulation. On top o f that, it is also important to analyse the restoration procedure when the network is exposed to different loads.
In this case, the aim is to observe the interactions am ong interleaving restoration procedures, hence the first version o f the simulation (fiill system implementation) was used. A s explained at the beginning o f Section 6.1, the first version does not allow the simulation o f very large system s with high request rates due to its memory cost. H ence, a small network w ith a restricted number o f users
was simulated. Instead o f increasing the number o f users or the request rate (that would cause the simulation run time to becom e unreasonably high), the loading effect w as emulated by decreasing
the processing capacity o f the network nodes - i.e. by increasing the time needed to perform a database operation. N ote tliat although decreasing the nodes processing capacity and increasing the
request rate are equivalent, they are not exactly the same and have slightly different consequences on the system. H ow ever, both cases represent a situation in which the nodes becom e increasingly unable to cope w ith the request rate.
Bv increasing the request rate, the restoration procedures start to seriously overlap with one another only when the delays begin to build up. By decreasing the nodes processing capacity, however, the restoration procedures can be made to overlap much before the system starts to show signs of breaking down. This is so because the restoration procedures take longer and longer to be completed and the probability that another one will be started before the previous one is finished gets increasingly high. For the purpose o f the simulation, this is a good point because this is the effect that needs to be observed - i.e. the interaction mnong concurrent procedures and their consequences on the system ’s performance. The input parameters for the plots in this section are listed in table 6.4.
N u m b e r o f lo catio n are as 4x4 = 16 location areas L o catio n a rea p o p u la tio n 500
fmove 0.6
tcall 0.4
A / 3.6x10* m sec
R e-co n fig u ratio n interv al 2.5x10* m sec
tp (link p ro p a g a tio n delay ) 1 m sec T im e to cross lo catio n area 1.0x10'" m sec
A verage call d u ra tio n 1.0x10'" m sec
D ata c o lle ctio n in terval 1.0x10^ m sec
T ab le 6.4: In p u t p a ra m e ters fo r p lo ts in Sectio n 6.2.3. 4.5EH05 - o 2.5F.H0 5 O 2 .0 E -0 5 I.0EH0 5 :> 5.0E-G4 0.0E-O0 3 0 0 0 too 200 4 0 0 5 0 0 n o d e s e r v i c e tim e ( m s e c ) - l o c a t i o n C h a n g e r e q u e s t s - c a l l r e q u e s t s - r e s t o r a t i o n p a c k e t s
Figure 6.27: R e -co n fig u ra tio n resu lts: variatio n o f av erag e en d -to -e n d d elay a cc o rd in g to node serv ice tim e.
Figure 6.27 plots tlie variation o f the average end-to-end delay for location change requests, call requests and restoration packets. As tlie node service time gets sufficiently high to start causing significant delays within the system, tlie average end-to-end delay grows at a much higher rate than for low service times. This is due to the time the packets are made to wait at the nodes input buffer. If the load had been raised by increasing the request rate, then tlie growth in the average end-to-end delay would be more gradual, suddenly shooting up as the system approaches breaking point (as it is the case for the plots o f Section 6.2.2). By increasing the node service time, on the
other hand, tlie growth in the average end-to-end delay is more accentuated, and the rate of increase becom es significimtly higher tlian tlie initial rate much before the breaking point.
2.0 p. «6 . 1.4 P 0 6 TO ^ l.OE-06 "g % 6.0 P 0 5 2 . 0 P H 0 5 0 . 0 P - O 0 200 3 0 0 4 0 0 6 0 0 0 100 500
n o d e s e r v i c e tim e ( m s e c ) -lo c a tio n change req u e sts
-call req u ests
Figure 6 .28; R e-co n fig u ratio n re.sults: av erag e en d -to -en d d e la y fo r re q u e sts that detected an error.
i . c iÛ ^
P
a. 0.6 Q 0, 0 .4 2%
0 2 I 0 100 200 3 0 0 4 0 0 5 0 0 6 0 0 n o d e s e r v i c e tim e ( m s e c )F igure 6.29: R c-co n fig u ra tio n resu lts: v ariatio n o f p e rc en tag e o f re q u e st p a ck e ts that d e te cte d an erro r w h ile b ein g tra v e llin g th e n e tw o rk a cc o rd in g to no d e serv ice tim e.
An interesting point to note from figure 6.27 is that the average end-to-end delay for restoration packets, initially higher than for location change and call requests becomes lower for very high values o f node service time. It is initially higher because the restoration packets are processed by tlie high level nodes between old and new parents that, due to the small topology being simulated, not only are re-directing restoration packets from multiple concurrent restoration procedures but also have a high probability o f being involved in a re-configuration procedure tliemselves. Hence these nodes are the most heavily loaded. Tliis causes the average end-to-end delay for restoration packets to becom e higher than for request packets.
A s the node serv ice time grows, however, the restoration procedure takes longer (see figure 6.30) and hence tlie availability o f nodes for rc-configuration becom es smaller, causing tlie re configuration rate to decrease (see figure 6.31). As less re-configurations are performed per unit time, a smaller number of restoration packets are issued, reducing the load on the re-configured nodes, and hence the delay in processing restoration packets. The reason why tliis is not reflected
on the requests average end-to-end delay is because as the node service time increases, the interleaving o f re-configuration procedures increases accordingly, causing inconsistencies to be generated. The request average end-to-end delay is reflecting the growing percentage o f packets that cause an error to be detected, consequently raising the average end-to-end delay. The average end-to-end delay for requests that encounter an inconsistency w hile travelling tlie network is plotted in figure 6.28. Figure 6,29 shows the growing percentage o f requests that detect an error.
The variation o f the restoration time (time required to restore the database after each re configuration) with increasing service time is plotted in figure 6.30. The curve's shape is determined by the average delay for restoration packets as plotted in figure 6.27.
% 3 .0 E O 6 0/ 2.0EH06 I.0E-G6 I . . . 0.0EÎ-O0 0 100 200 3 0 0 4 0 0 5 0 0 6 0 0 n o d e s e r v i c e tim e ( m s e c )
Fig u re 6.30: R e-co n fig u ra tio n resu lts: av erag e re sto ra tio n tim e plo tted a g ain st node serv ice tim e.
4.5E-07 4.0 E -0 7 2 2.5E-0 7 2 .0 E -0 7 9 l.OE-0 7 5.0E-0 8 O.OE-CO 0 100 200 3 0 0 4 0 0 5 0 0 6 0 0 n o d e s e r v i c e lim e ( m s e c )
Figure 6.3 1 : R e-co n fig u ra tio n resu lts: av erag e rc -c o n fig u ratio n rate p lo tted ag ain st n o d e serv ice tim e.
As the restoration procedure takes increasingly longer to be completed, the number o f nodes available for re-configuration at any given time gets increasingly lower, causing tlie re configuration rate to fall simply because there are no nodes available to take part in a re configuration. This IS shown in figure 6.31. The simulation has been designed so tliat while a node
is acting as an old parent or a reconnected child, then it cannot participate in another reconfiguration either as old parent or as a reconnected child Note that no restrictions are placed on new parents, since they do not change their operation during the restoration procedure.
1.0 l i « 6 % O 6.0124] 5 5 O 4.0EH05 2.0EH05 0.0EH00 DO 200 300 400 500 600 0
node service linie (m se c)
Figure 6 .3 2 ; R e -co n fig u ra tio n re.sults: v ariatio n o f tim e interval betw een e iro r d etectio n and co rre c tio n a c c o rd in g to no d e serv ice tim e.
As mentioned above, the interactions among concurrent restoration procedures are likely to cause inconsistencies to anse. Those inconsistencies are detected when the corresponding infonnation is required, causing a recover) procedure to be performed (as described in Section 4 . 1). As the delays within the system start to build up, the time inteiwal between error detection and correction increases accordingly (see figure 6.32). When the delay between detection and correction becomes too high for the system to cope, tlie correction rate falls below the value of the detection rate. This can be seen from figure 6.33
3.0E -0 5 r — $ o 2 .0 E -0 5 ^ ^ 1.5E-0 5 <0 t O I.OE-0 5 ■è 5.0 E-06 O £ O.OE-00 DO 200 300 4 0 0 500 6 0 0 0
n o d e s e r v i c e tim e ( m s e c ) - d e te c tio n rate
- c o r r e c tio n rate
Fig u re 6.33: R e-co n tig u ra tio n resu lts: erro r d etectio n a n d co rre c tio n rates v ersu s no d e serv ice tim e.
Figure 6.34 gives the variation o f tlie average load on each node according to their initial hierarchical level for increasing node service times. Note that the most heavily loaded nodes are the ones that are initially at the first hierarchical level. Those are the nodes that are most exposed
to the rc-connguration process because they can becom e reconnected children, old parents or new parents and they are the nodes most likely to be in the path between old parent and reconnected child forcing them to re-direct most o f the restoration packets.
80 70 c 4 0 K) 0 100 200 3 0 0 4 0 0 5 0 0 6 0 0 0 n o d e s e r v i c e tim e ( m s e c )
- local exchanges . rirs: level nod es
I'ig u re 6.34: I^e-co n fig u ratio n results: load v ariatio n (g iv e n as percen tag e o f no d e c ap a c ity ) a cc o rd in g to each no d e initial hierarch ical level.
The root node is not allowed to reconnect, hence the two roles it can perform are either o f old parent or new parent. Due to tlie small simulated topology, the root is also forced to redirect most o f the restoration packets on their way from reconnected child to old parent. Tlie least affected nodes are the local exchanges that can only assume the role o f reconnected children. The local exchanges have only to process the restoration packets generated due to tlieir own re configurations. The nodes are also significantly loaded by the recover) mechanisms performed to correct the inconsistencies generated due to the interleaving of concurrent restoration procedures. Tlie root node is particularly affected by this factor because it has to process a flood fill packet and a rebuilding packet for each o f the detected inconsistencies.
100 l/i S a-
I
0i
0 100 200 3 0 0 4 0 0 5 0 0 6 0 0 n o d e s e r v i c e t im e ( m s e c )Figure 6 .35: R e-co n fig u ra tio n resu lts: a v erag e p e rc en tag e o f d elay ed p ackets v ersu s no d e serv ice tim e.
Figure 6.35 shows llie average percentage o f packets that suffer delay at som e point in their paths. N ote that the system approaches the breaking point because o f the inconsistencies caused by the interleaving o f concurrent restoration procedures and because o f the increasing probability o f clashes am ong tlie ordinary request transactions and the recovery m echanisms in place and not because o f the extra signalling load generated by the re-configurations. In fact tlie signalling load directly generated by the re-configurations decreases as the delays start to build up. This occurs because the number o f nodes available to re-configure decreases as the delays start to grow,
causing the re-configuration rate to fall.
The results presented in this section have demonstrated the feasibility o f the management re configuration m echanism , w hich is sufficiently robust to survive um ealistically high delays within the system. The m anagement re-configuration mechanisms together with the recovery procedures make the system robust and flexible, capable o f adapting to an evolving environment.
6.3 Sum m ary
In order to evaluate the system's performance and to study the introduction o f tlie recovery mechanism s, the system was simulated for the particular application o f personal m obility service. A brief description o f the simulation design is given in Section 6.1. The m odels used in the two versions o f the simulation, the system ’s full implementation and the system ’s statistical model, are explained in detail. The obtained results and relevant discussions are presented in Section 6.2. The first set o f results demonstrates the system's scalability, by show ing the weak dependency on system ’s size o f the sw itch throughput and packets average end-to-end delay. The second set shows the feasibility o f the adopted recovery mechanism s and how the system responds to different error rates. The third and final set o f results demonstrates the feasibility o f the re configuration procedure and how the system supports re-configuration under different loading conditions.
The system is shown to be able to cope with high error rates and capable o f performing re configurations under high loading conditions, without interruption to the system ’s normal operation and with limited effect on the system ’s performance.