Brainy: Effective Selection of Data Structures

runmidgeAI and Robotics

Oct 20, 2013 (3 years and 9 months ago)

73 views

Brainy:

Effective Selection of
Data Structures


Why data structure selection is important


How to choose the best data structure for a
specific application


Related work


Brainy


Evaluation


Future work


S
ummary

Niklaus

Wirth

Vector (Store
i
th

item in A[
i
])


Alternative Data Structures

Singly or Double
linked list


Find the
i
th

item

O(1)

O(n)

Insert
or delete an
element

O(n)

O(1)


less cache misses,

but may move all
elements when dynamic
adjust size

Memory

accessing

more cache misses


Why data structure selection is important


How to choose the best data structure for a
specific application


Related work


Brainy


Evaluation


Future work


Summary



Data structure library




Asymptotic analysis

Developer depend
on

DSLs are designed for common case,

n
ot consider application
-
specific situation

Asymptotic analysis is not accurate to data
structures. Better asymptotic behavior may
perform worse (for example, Fibonacci heap);
Same asymptotic behaviors with different
performances (splay tree often perform better
than other balanced binary search trees).



Which interface functions are invoked, (how often
the functions are used…)


What is the input for the data structure (data
elements size, the number of elements…)


Which kind of underlying architecture the
program running (cache size…)

What we need take into account?

The paper shows, between two different architectures Intel Core
Q6600 and Intel Atom N270, 43% of the randomly generated
application have different optimal data structures.

Related Work

L.Liu

and
S.Rus

Perflint
: A Context Sensitive

Performance Advisor for C++ Program

O.Shacham
,
M.Vechev
, and
E.Yahav
. Chameleon:

adaptive selection of collections


Why data structure selection is important


How to choose the best data structure for a
specific application


Related work


Brainy


Evaluation


Future work


Summary


How to get the training examples?

Is using real applications to train the
machine learning algorithm a good suggestion?

How to get the training examples?

Condition:

In order to avoid over fitting,
e
very portion of the design
space (input/ interface function) must be fully represented.

Application Generator: creating a variety
of applications that test different parts of
the overall space.


Feature Selection (Using genetic algorithm)


Take a given subset of features as a
chromosome, and get the global optimum by
mutation.


Artificial Neural Network


Why data structure selection is important


How to choose the best data structure for a
specific application


Related work


Brainy


Evaluation


Future work


S
ummary


Performance improvement Brainy achieved

Evaluation

27% on Core2 and 33% on Atom


Why data structure selection is important


How to choose the best data structure for a
specific application


Related work


Brainy


Evaluation


Future work


S
ummary



Exclude the interference from intervening
instructions

Future Work


Why data structure selection is important


How to choose the best data structure for a
specific application


Brainy


Evaluation


Future work


S
ummary



Selection of Data Structures
is important


Using
machine learning in data structure
selection


Hardware features


Reduced overhead

Summary