31

after receiving the reply from the Web server, it waits for a certain interval based on the

Pareto distribution, before it generates the next request.

Table 3.3. Distribution Models Used in Simulation

Parameter

Description

Value

Incoming request interval

Pareto distribution

P (x) =  k

/x

 +1

,   = 1.1

-(ln x- )2

Distribution of objects

Lognormal distribution

P (x) =

1

  e

2 2

,

x  2 

(body)

  = 1.16,   = 0.9, x   10

Distribution of objects

Pareto distribution

P (x) =  k

/((1 - (

k 

p

)

)   x

 +1

)

(tail)

,x > k

, k

= 10 p = 2000,   = 1.1

3.4.2

System Validation

To validate our simulator, we have measured the round trip time in a 16 node

Linux cluster to compare with the simulator results. Figure 3.5 shows the latency com 

parisons from the simulator and the measurements. To test the round trip latency, we

used a pingpong application that exchanges a fixed size of data between two nodes, be 

cause the pingpong test represents the network latency correctly without incurring the

system overhead. In addition, we vary the number of pingpong processes by 1, 2 and 4.

In Figure 3.5, we show two latency results since VIA supports two communication modes:

non blocking and blocking. In both the Figures 3.5, the results from the simulator match

well with these from the real experiments.