Database Reference
In-Depth Information
Figure 8. System size planning analysis results
Replication Degrees (PARD) and Availability
Analysis (AA).
average runtime advantage of 50 nodes is very
small when compared to 25 nodes, due to the
increase in transfer and merge overheads among
so many nodes.
System Size Planning (SSzP)
Performance Analysis of
Replication Degrees (PARD)
Our first experiment ran a System Size Planning
analysis with ChunkSim, also repeating the same
experiment in our lab setup to compare the estima-
tion with actual runtimes. For these experiments
the placement was homogeneous and the replica-
tion degree was 1 (one copy of each chunk). The
results are shown in Figure 8. Since there were
only 16 PCs in the physical setup and we wanted
to include average runtime estimates for larger
system configurations as well (25 and 50 nodes
in the Figure), we assumed performance indexes
for additional nodes beyond 16 using the same set
of those 16 nodes in a round-robin fashion.
The results from Figure 8 show that the error
in the ChunkSim average runtime estimation was
within a 10% error from the actual runtimes for
those combinations that were tested against both
the simulator and the system (those with less than
16 nodes). It also allows the user to see what the
average query runtime will be depending on the
number of nodes. For instance, with 5 nodes the
average runtime will be 695 secs, while with 25
nodes it will be 134 secs. We can also see that the
The PARD experiment allows us to analyze how
different replication degrees affect runtime. For
this experiment we considered homogeneous
placement and varied the replication degree on 16
nodes. Figure 9 shows the results. The replication
degree is depicted in the x-axis and represented as
a fraction of the number of original chunks. There
are two runtime interval boundings - the homo-
geneous (PL-H) and Full Mirroring (FM) curves
(these concern only Local Processing) - and three
more runtimes: the Local Processing estimation
and the Query Runtime estimation are outputs of
ChunkSim, while theActual Query Runtime is the
actual time taken in the lab experimental setup.
From Figure 9 we can see that ChunkSim does
a good job estimating the runtime, and we can also
see how the runtime evolves as the replication
degree increases. One replica improves average
query runtime by about 13%, and 2 replicas im-
prove by about 23%, both when compared with
Search WWH ::




Custom Search