Database Reference
In-Depth Information
1
1
1
JDA
M
GrePP
JDA
M
GrePP
Linear Plan
Bushy Plan
0.8
0.8
0.8
0.6
0.6
0.6
0.4
0.4
0.4
0.2
0.2
0.2
0
0
0
1.5x
3x
5x
1.5x
3x
5x
1.5x
3x
5x
Freshness
P
redicate
(as a Multiple of Window Predicate)
Freshness
P
redicate
(as a Multiple of Window Predicate)
Freshness Predicate
(as a Multiple of Window Predicate)
(a) Linear
(b) Bushy
(c) Resources Used for freshness
Fig. 14.
Evaluation of freshness predicates
more significantly than the change in non-root operators. The performance of the JDA
M
strategy is significantly lower than JAQPOT. Figure 12.b illustrates the results for the
bushy
plan with changes in the selectivities at 3 and 7 mins, just like the linear plan. Here
JAQPOT again produces high throughput while JDA
M
continues to produce output at a
very low rate.
Impact of Changing Available Resources.
These charts (Fig. 13) summarize the per-
formance of the techniques over the complete range of
values from 0% to 100% of
saturation. A variety of parameter settings are used, as in Table 11. The charts depict
μ
μ
as a percentage of the
saturation resources
. On the y-axis, the throughput produced by
JAQPOT and JDA
M
strategies, averaged over several runs, is shown as a percentage of
the saturation throughput.
For the linear plan (Fig. 13.a) JAQPOT outperforms the JDA
M
strategy producing
more than 80% of the saturation throughput while using only 60% of the resources.
Averaged over the different
μ
values, JAQPOT consistently beats the JDA
M
strategy
by producing 3
×
as many tuples/min on average, with a maximum of 6.5
×
at 40% of
saturation resources.
For the bushy plan (Fig. 13.b), the trends are similar. Overall, JAQPOT performs
much better for the linear plan than the bushy plan. In the linear plan, all the
staleness
susceptible states
can be synchronously refreshed as they are covered by fewer (pos-
sibly single) path, whereas, for the bushy plan, atleast two paths require resources for
freshness
fulfillment, thus leaving fewer resources for throughput optimization.
5.2
Evaluating Result Freshness
The purpose of these experiments is twofold, namely, (a) to establish that result stale-
ness is aggravated by the JDA approaches, and (b) to analyse how much of the resources
are spent in satisfying
freshness
. The staleness of results is measured by counting the
number of tuples produced that violate a given
freshness
predicate
I
as per Def. 1.
Result Staleness in Join Adaptation.
Next, we substantiate our hypothesis that the
throughput optimizing schemes aggravate the result staleness problem
. We compare
the JDA
M
and the GrePP knapsack solver omitting the GH-WMSCP component such
that GrePP produces stale tuples in the absence of GH-WMSCP. We perform these
experiments on the linear and bushy plans (Figs. 9.b and c). For each plan, we create
many scenarios by varying the parameter settings (Table 11). Here,
F
μ
is fixed at 300.
