Information Technology Reference
In-Depth Information
of nodes per PG gives results with the lowest message usage. This is
mostly due to the fact that in both the remaining cases the most likely
nodes to fail were grouped together creating highly unstable PGs. Not
only will this increase the updating activity within the PG but all of
their TEC contacts have to update their tables too, creating a constant
surge of trafi c in these locations. Despite the creation of highly active/
unreliable PGs, (churn) observations show that the network still proves
to be stable and robust, which is further verii ed by the absence of bursts
in usage.
13.6
This chapter has presented PIndex as a viable peer-to-peer-based solu-
tion for grid information services, which builds on top of MDS4. Colored
Petri nets were used to create a simulation model that could simulate
many thousands of nodes in the PIndex network. Having simulated
network populations of up to 10,000 computing nodes and investigated
a variety of peer group grouping methods, the simulated results show
PIndex to be highly salable and stable in addition to handling node
failures.
Having discovered that PIndex remains stable despite the grouping of
computing nodes with high failure rates, and that an expected rise of
message usage and node usage has respective peak gradients of 1 and
0.78, this proves that PIndex is a highly scalable network structure and a
viable solution to the grid information service problem. However, hav-
ing modeled PIndex using CPN that included Poisson distribution rates
and normal distribution probabilities, these can only give an indication
of the performance PIndex will have under these conditions. The CPN
model did not cover the underlying network infrastructure that PIndex
was running on and at best related to a local network coni guration.
However, given the results the following two improvements to PIndex
could be made.
Having already stated that the underlying network was not included in
the CPN model, an additional improvement that could be made to PIndex
is to include proximity in the PIndex algorithm such that the geographi-
cally local computing nodes will belong to one PG.
Although in the simulated results we observe a startup l ux, this levels
off to a stable state. On closer inspection oscillations are seen that can
only be explained by PGs updating at regular intervals, which creates
variations in resource usage. By implementing a time slotting mechanism
for PG updates, resources can be effectively utilized.
Conclusion
Search WWH ::
Custom Search