Database Reference
In-Depth Information
5
Experimental Evaluation
We now examine the effectiveness of JAQPOT compared against the state-of-the-art
JDA technique (described in Section 3). We implemented JAQPOT and JDA M within
the CAPE stream engine [15].
Parameter Va lue
Arrival rates ( λ i ) 300 1200 tuples/sec
Window sizes of states ( | S I | ) 200 5000 tuples
Join selectivities (
σ i )
0.01
0.1
Available Resources ( μ )
0 100 % of saturation
I )
Freshness predicates ( F
1.5 ×∼ 5 × of | S I |
Fig. 11. Experimental parameters with values
Objectives. The goal of our experimental study is to substantiate the observations of our
analytical study (Sec. 4) that JAQPOT is capable of producing near optimal through-
put together with maintaining result freshness . We evaluate JAQPOT and its competitor
JDA M policy by measuring their performance in (a) producing query throughput ,and
(b) fulfilling the freshness predicates . We examine the following critical questions with
focus on the two performance measures:
i , and S I as well as
the query shape on the throughput produced by the JDA techniques?
What is the impact of stream and query parameters, such as
λ
i ,
σ
How does the throughput produced by JAQPOT and JDA M techniques compare with
the saturation throughput 8 with change in
μ
?
In the absence of the set-coverage solver (GH-WMSCP), how badly do the throughput
optimization techniques perform in terms of the result freshness?
In JAQPOT, what fraction of resources get assigned for fulfilling freshness as opposed
to achieving high throughput?
Experimental Setup. All experiments are run on a machine with Java 1.6 runtime,
Windows 7 with Intel(R) Core(TM)2 Duo CPU@2.13 GHz processor and 4 GB RAM.
All techniques are tested rigorously using synthetic streams and distinct query shapes
with arbitrary parameter settings (Table 11). Further, the applicability to a real-world
application is also verified using the Weatherboards data set [1]. The results for the
experiments with the real data set are not included in this paper. Please refer to technical
report [12] for details.
5.1
Throughput Production in Synthetic Data
The goal of our experiments is to compare the throughput produced by both the JDA
techniques under (a) fluctuating streams, and (b) changing resource availabilities. We
8
The minimum total resources required to process the full query workload with no CPU limita-
tion are called the saturation resources . The corresponding throughput produced is called the
saturation throughput .
 
Search WWH ::




Custom Search