V. NUMERICAL RESULTS The numerical results are presented in this section to quantify the performance of the CMV algorithm. The characteristics of RPs in the Gnutella system are measured in [2] and [17]. The results in [17] show that about 60% peers have 0.2 or less online probability, and 10% peers have more than 0.8 online probability; Moreover, about 50% peers have 60 minutes or less and 10% peers have 300 minutes or more in each online session. The IP address of a peer may be changed for each rejoining; the results in [2] indicate that about 40% peers changed their IP address in one day and about 50% in seven days. We select the parameters based on these observations, our parameters are set as: Pon l = 0.2, Pon h = 0.4, Tc l = 270 (min), Tc h = 300 (min), λl = 0.05 (min−1) and λh = 0.2 (min−1). s is set as 10. The parameters of the VPRs are set as Tc v = 330 (min−1), λv = 0.5 (min−1). Two different file update arrival rates (α = 0.002 and 0.2 min−1) are used, both LRPs and HRPs have the same file update arrival rate in their online time. The optimal number of VRPs (Nv) varies with different system conditions, we assume all updates are generated by HRPs and LRPs to maintain a constant file update rate (i.e., α = NhαhPon h + NlαlPon l = constant) in the whole system. We set Cbrd = 1, 000, 000 and Cfld = 10, 000 messages. A flooding search costs fewer messages because multiple VRPs may be online with changed IP address. Moreover, an online HRP may be searched and hence a more recent information of the VS can be obtained, thus resulting in reduced search cost. We use the number of overhead messages per query (OHPQ) and the number of retrieved files per query (FPQ) as two metrics. OHPQ is defined as H/λ, and FPQ as F/λ. The parameter setup ensures that the file maintenance cost of an HRP and LRP is minimized. We set PcIP v = 0.3, Nh = 500 and Nl = 4000. Figure 2 shows optimal Nv as a function of Pon v . We observe that the optimal value of Nv decreases from about 33 toapproximately 10 as Pon v increases from 0.4 to 0.9. This is due to the fact that a VS composed of smaller number of VRPs with larger Pon v can provide the same availability as a VS composed of larger number of VRPs with smaller Pon v . The file update rate has very little impact on the optimal Nv in this case. This is because the total update rate is much smaller than the total file access rate. This is true in many database and file sharing systems. From Figure 3, we can see that the FPQ is only dependent on the file update rate. A fast updating file corresponds to a larger FPQ, because the replicas of a fast updating file become stale quickly, subsequently, the file accesses from these replicas need to be retrieved from the VS, thus resulting in a larger FPQ. The OHPQ is decreased from approximately 2.5 to just above 1 as Pon v increases from 0.4 to 0.9. A large Pon v leads to a small optimal Nv and then small OHPQ. The results indicate that the overhead messages for file maintenance are very low in the CMV algorithm, especially for the VS composed of high available RPs.
zax1987 (04/28/2008)
V. NUMERICAL RESULTS
The numerical results are presented in this section to quantify
the performance of the CMV algorithm. The characteristics
of RPs in the Gnutella system are measured in [2] and
[17]. The results in [17] show that about 60% peers have 0.2 or
less online probability, and 10% peers have more than 0.8 online
probability; Moreover, about 50% peers have 60 minutes
or less and 10% peers have 300 minutes or more in each online
session. The IP address of a peer may be changed for each rejoining;
the results in [2] indicate that about 40% peers changed
their IP address in one day and about 50% in seven days. We
select the parameters based on these observations, our parameters
are set as: Pon
l = 0.2, Pon
h = 0.4, Tc
l = 270 (min),
Tc
h = 300 (min), λl = 0.05 (min−1) and λh = 0.2 (min−1).
s is set as 10. The parameters of the VPRs are set as Tc
v = 330
(min−1), λv = 0.5 (min−1). Two different file update arrival
rates (α = 0.002 and 0.2 min−1) are used, both LRPs
and HRPs have the same file update arrival rate in their online
time.
The optimal number of VRPs (Nv) varies with different system
conditions, we assume all updates are generated by HRPs
and LRPs to maintain a constant file update rate (i.e., α =
NhαhPon
h + NlαlPon
l = constant) in the whole system. We set
Cbrd = 1, 000, 000 and Cfld = 10, 000 messages. A flooding
search costs fewer messages because multiple VRPs may
be online with changed IP address. Moreover, an online HRP
may be searched and hence a more recent information of the
VS can be obtained, thus resulting in reduced search cost.
We use the number of overhead messages per query (OHPQ)
and the number of retrieved files per query (FPQ) as two metrics.
OHPQ is defined as H/λ, and FPQ as F/λ. The parameter
setup ensures that the file maintenance cost of an HRP
and LRP is minimized. We set PcIP
v = 0.3, Nh = 500 and
Nl = 4000.
Figure 2 shows optimal Nv as a function of Pon
v . We observe
that the optimal value of Nv decreases from about 33 toapproximately 10 as Pon
v increases from 0.4 to 0.9. This is due
to the fact that a VS composed of smaller number of VRPs with
larger Pon
v can provide the same availability as a VS composed
of larger number of VRPs with smaller Pon
v . The file update
rate has very little impact on the optimal Nv in this case. This
is because the total update rate is much smaller than the total
file access rate. This is true in many database and file sharing
systems.
From Figure 3, we can see that the FPQ is only dependent
on the file update rate. A fast updating file corresponds to a
larger FPQ, because the replicas of a fast updating file become
stale quickly, subsequently, the file accesses from these replicas
need to be retrieved from the VS, thus resulting in a larger FPQ.
The OHPQ is decreased from approximately 2.5 to just above
1 as Pon
v increases from 0.4 to 0.9. A large Pon
v leads to a
small optimal Nv and then small OHPQ. The results indicate
that the overhead messages for file maintenance are very low
in the CMV algorithm, especially for the VS composed of high
available RPs.
cincinnati (04/28/2008)
其实我觉得你可以先将自己翻译的东西粘上来，然后大家一起帮着改，因为这么专业的东西，想必还是你自己最有权威吧！