Figure 3a shows the time required for multimedia
processing and event decision. For multimedia
processing, the delay is less than 10 seconds for
data smaller than 0.6 Kbytes. However, as the data
size increases, the delay grows dramatically. This
feature comes from the MapReduce framework, in
which most time is spent shuffling data. For event
decision, the delay increases slowly as the data size
increases, because the parallel data processing consumes
time in calculation without exchanging data.
The sensor data decision application takes a slightly
smaller workload than the multimedia analysis
when processing the same size of data with a small
request pressure (less than 0.6 Kbytes). However, as
the number of requests further increases, the multimedia
analysis shows a more sensitive growth than
the data decision, because the multimedia analysis
includes a larger workload of feature extraction on
real-time data, which suffers from the mass of data
shuffling between distributing nodes.
Figure 3a shows the time required for multimediaprocessing and event decision. For multimediaprocessing, the delay is less than 10 seconds fordata smaller than 0.6 Kbytes. However, as the datasize increases, the delay grows dramatically. Thisfeature comes from the MapReduce framework, inwhich most time is spent shuffling data. For eventdecision, the delay increases slowly as the data sizeincreases, because the parallel data processing consumestime in calculation without exchanging data.The sensor data decision application takes a slightlysmaller workload than the multimedia analysiswhen processing the same size of data with a smallrequest pressure (less than 0.6 Kbytes). However, asthe number of requests further increases, the multimediaanalysis shows a more sensitive growth thanthe data decision, because the multimedia analysisincludes a larger workload of feature extraction onreal-time data, which suffers from the mass of datashuffling between distributing nodes.
การแปล กรุณารอสักครู่..
