Database Reference
In-Depth Information
If the mapper Map1 is responsible for processing the packet number 1 or
P
1, it sends a message
miniChunkRequest
(1
,pack size
)to
m −
1mappers
(assume
m
=4).Eachmapper
Map
j
(
j
in [2
..n
]) answers with a message
SendminiChunk
(
a
j
,R
[
j,
1
,x
]), where
a
j
is a header of the chunk j that is in-
deed the
j
th
vector of the key matrix
A
and
R
[
j,
1
,x
] is the vector composed of
x
elements starting from the first element in the chunk received by the
Map
j
mapper (see Figure 2).
In classic MapReduce frameworks, there is no communication between map-
pers.
3.4 Extraction Phase (Combine)
Once the data is collected, each mapper first determines the inverse key matrix
from the
m
keys that it possesses (its own and those it had just received in
the collect phase). Thereafter, by multiplying the inverse matrix by the vector
composed of the elements collected, each mapper gets the packet to perform the
map task on it. These instructions match the treatments performed by the IDA
Combine function. Figure 3 summarizes the steps of this phase carried out by
the mapper Map1.
Fig. 3.
Extraction of data by Map1
Comparing our approach to classical MapReduce systems, the preparation
and scatter phases are the same. Our preparation phase may take longer time,
because it does not only split data but apply a more complex procedure. Then
classically, workers start immediately their map tasks, while in our approach,
each mapper needs extra treatment to clarify its data. The procedure is close to
deciphering with a shared key. Mapper must collect the key to decrypt its data
in the extraction phase.
Search WWH ::
Custom Search