ChunkSim - Data Warehousing Design and Advanced Engineering Applications - page 147

Figure 9. Performance analysis for replication degrees results

0.1 replicas. Beyond 2 replicas the improvement
is small, at a significant increase in overall dataset
size and therefore of loading and storage costs.
This seems to indicate that 2 replicas would be an
interesting option for this particular layout.

The observations from the previous experi-
ments apply for the setting that was tested, con-
cerning both dataset and chunk configurations,
workload properties and performance indexes
of the 16 nodes. With other settings, ChunkSim
would allow us to reach conclusions and plan
for the corresponding configurations as well. As
a result, ChunkSim is an adequate tool to help
plan, analyze and decide on the number of nodes,
placement and replication layouts and replication
degrees, depending on requirements concerning
performance (average response times) and avail-
ability (tolerance to nodes failing or being offline)
objectives.

Availability Analysis (AA)

The next experiment predicts what happens when
nodes go offline in our 16 node experimental setup
using ChunkSim. Figure 10 shows the results.
The point where each curve starts in Figure 10
indicates the minimum replication degree that is
necessary in order to tolerate a certain number of
node failures (we considered tolerance whenever
the Failure Rate of experiment runs was below
10%).
Figure 10 indicates that while it is necessary
to have at least 1 and 2 replicas to tolerate 1 and
2 node failures respectively, for this configuration
3, 4 and 5 node failures can be tolerated with 2,
3 and 3 replicas respectively. These experiment
results also show a steep curve when 1 or 2 nodes
fail with a replica degree less than 3. A similar
behavior can be observed when 3 nodes fail with
less than 4 replicas, when 4 nodes fail with less
than 5 replicas and when 5 nodes fail with less
than 6 replicas.

CONCLUSIONS AND FUTURE WORK

In this paper we have proposed ChunkSim, an
event-based simulator for analysis of load and
availability balancing in chunk-wise paral-
lel data warehouses. We discussed first how a
shared-nothing system can store and process a
data warehouse chunk-wise, and use an efficient
on-demand processing approach. Then we dis-
cussed ChunkSim model parameters, the set of
parameters that the simulator must collect to be
able to simulate execution. We also discussed