Prototype Based Evaluation

In order to validate the results obtained with simulation, the performance of the sequencer based and statistically estimated optimistic total order protocols imple- mentation of Section 4.3 have been evaluated.

Measurements presented below were obtained by running the protocol implement- ations on top of the simulation infrastructure described in Section 5.4.2.

4.4. PROTOCOL EVALUATION 65 15ms 15ms 15ms 15ms 15ms H0 H4 H3 H2 H1 0,12ms 0,12ms 0,12ms 0,12ms 0,12ms (a) Star. 10ms 10ms 10ms 10ms 10ms H0 H4 H3 H2 H1 0,12ms 0,12ms 0,12ms 0,12ms 0,12ms 10ms 10ms 10ms 10ms 10ms (b) Ring. 10ms 10ms 10ms 10ms 0,12ms H1 10ms H0 0,12ms 10ms H4 0,12ms 10ms H3 0,12ms 10ms H2 0,12ms 10ms (c) Bus.

Figure 4.15: Network topologies.

The protocols were evaluated in the 3 network topologies depicted in Figure 4.15. All the experiences were conducted with 5 processes sending messages at a sim- ilar data rate. The star topology (Figure 4.15a) is a symmetric network composed of a central router to which all the leaf networks connect through a 15 ms WAN link. In the backbone network of the ring topology (Figure 4.15b), each router is connected to two adjacent routers, through a 10 ms WAN link, forming a ring. Each of the leaf networks is connected to one of those routers through a 10 ms WAN link. In the bus topology (Figure 4.15c), the backbone network is composed of 3 inner routers and 2 outer routers. Each outer router connects to one of the inner routers, and the inner routers connect to two routers of the backbone net- work. Each connection as a 10 ms latency. As in the ring topology, each of the leaf networks is connected to one of the backbone routers through a 10 ms WAN link. Common to all topologies is the fact that adjacent processes are separated by a latency of 30 ms.

In the experiences, message inter-arrival for each process is obtained from a neg- ative exponential distribution with the corresponding parameter to accomplish the desired message rate. Each process records the instant it sends and receives each message optimistically and authoritatively. After finishing the experience, those records are processed and the value of the spontaneous total order calculated. The

0 20 40 60 80 100 0 10 20 30 40 50 60 70 80 90 100

% spontaneous total order

per process message rate/s Statistically estimated TO

Sequencer based TO

(a) Start topology.

0 20 40 60 80 100 0 10 20 30 40 50 60 70 80 90 100

% spontaneous total order

per process message rate/s Statistically estimated TO Sequencer based TO (b) Ring topology. 0 20 40 60 80 100 0 10 20 30 40 50 60 70 80 90 100

% spontaneous total order

per process message rate/s Statistically estimated TO

Sequencer based TO

(c) Bus topology.

Figure 4.16: Spontaneous total order.

values of spontaneous total order for each topology are depicted in Figure 4.16. The star topology resembles network topology 1 of the simulation model where the latency between every two processes is 30 ms. As predicted in the simulation (Figure 4.11a), the optimistic protocol presents a much higher spontaneous total order, being higher than 90% for message rates up to 100 messages per second per process, as depicted in Figure 4.16a. The spontaneous total order of the op- timistic protocol observed in the Ring topology, and depicted in Figure 4.16b is lower than in the Star topology, but higher than the total order protocol. Even in the Bus topology, depicted in the Figure 4.16c the optimistic protocol present better spontaneous total order than the total order protocol. In this case the values became similar for message rates higher than 80 messages per process.

The gain in spontaneous total order depends on the topology being considered and on the message rate. Figure 4.17 presents, for each of the topologies, the ratio of spontaneous total order between the statistically estimated total order protocol and the sequencer based total order protocol. It can be seen that the statistically estimated total order protocol presents better results than the sequencer based total order algorithm, and in several situations it presents spontaneous total order values 3 to 4 times higher than the total order protocol.

4.4. PROTOCOL EVALUATION 67 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 0 10 20 30 40 50 60 70 80 90 100

Statistically estimated TO/Sequencer based TO

per process message rate/s

(a) Star topology.

0 0.5 1 1.5 2 2.5 3 3.5 4 0 10 20 30 40 50 60 70 80 90 100

Statistically estimated TO/Sequencer based TO

per process message rate/s

(b) Ring topology. 0 0.5 1 1.5 2 2.5 3 3.5 0 10 20 30 40 50 60 70 80 90 100

Statistically estimated TO/Sequencer based TO

per process message rate/s

(c) Bus topology.

Figure 4.17: Comparison of spontaneous total order.

0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80 90 100 Optimistic window

per process message rate/s

Sequencer based TO Statistically estimated TO

(a) Star topology.

0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80 90 100 Optimistic window

per process message rate/s

Sequencer based TO Statistically estimated TO (b) Ring topology. 0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80 90 100 Optimistic window

per process message rate/s

Sequencer based TO Statistically estimated TO

(c) Bus topology.

0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80 90 100 Delivery window

per process message rate/s

Sequencer based TO Statistically estimated TO

(a) Star topology.

0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80 90 100 Delivery window

per process message rate/s

Sequencer based TO Statistically estimated TO (b) Ring topology. 0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80 90 100 Delivery window

per process message rate/s

Sequencer based TO Statistically estimated TO

(c) Bus topology.

Figure 4.19: Delivery window.

protocols in each topology. From the figures it can be observed a small decrease in the optimistic window and an increase in the protocol latency, as expected from the results of the simulation model.

As in the simulation the results demonstrate that the higher degree of spontaneous total order of the optimistic protocol is not obtained sacrificing the optimistic win- dow nor the protocol latency. With respect to the optimistic window, as expected it is smaller as it is being used to improve the spontaneous total order. The pro- tocol latency suffers from the fact that the sequencer is delaying the ordering of its own messages. Being so, there is a small increase in the protocol latency.