Information Technology Reference
In-Depth Information
time compared to the HP-GPP (Intel i7-2600). As it is shown in this figure the
execution time of the embedded processor is 3x to 8.4x higher than the execution time
of the HP-GPP while it is 3x to 5x higher than the low power GPP. The lower
execution time of the GPPs can be justified by the higher clock frequency and the
more advanced instruction set (e.g. deeper pipeline scheme, larger L3 cache and more
advanced branch prediction schemes). The highest difference in the execution time is
however noticed on the
matrix multiplication
which is not used in the cloud
computing applications. In the commonly used cloud computing tasks such as the
word count
,
histogram, linear regression
and
string match
the execution time of the
embedded processor range from 3x to 5x higher than the HP-GPP.
Normalized Execution Time
HP-GPP
LP-GPP
EP
10
9
8
Cloud Applications
7
6
5
4
3
2
1
0
Histogram
Linear_regr
String match
Word_count
Matrix_mult
Application
Fig. 3.
Normalized Execution time for difference applications
The higher speedup of the
matrix multiplication
can be also justified by Figure 4. This
figure shows the average miss rate of the branch predictions for the Intel low power
processor. In this figure it is clear that the typical benchmark applications such as the
matrix multiplication are much more predictable due to the control structure and
therefore the branch miss rate is much smaller than the common cloud application tasks.
Branch Prediction Miss Rate
5
4
3
2
1
0
Histogram
Linear_regr String_match Word_count Matrix_mult
Applications
Fig. 4.
Branch prediction miss rate
