Jacobi Relaxation Term Project Plan

monkeybeetleΛογισμικό & κατασκευή λογ/κού

2 Δεκ 2013 (πριν από 3 χρόνια και 15 μέρες)

59 εμφανίσεις

CS
410 GPU
/
Term project
/
Hybrid CUDA

Team


Name

Email

Login

Tochi

tochi.ece@gmail.com

nwaoduht

David

Davidhsmith255@gmail.com

dsmit2

Andy

ajost@synopsys.com

Josta

Weixiong

luwwww@gmail.com

Luw

Matthaus

mlitteken@gmail.com

matthaus





Jacobi Rela
xat
ion

T
erm
P
roject

Plan




Abstract


A customer
wants to

build a
Jacobi Relaxation

application for
computing the voltage
potential of electrostatics area.

The customer has made
it
clear th
at the minimum
requirement is to calculate the Jacobi relaxation with
a
n

exceptional speed



calculation
throughput which

will rely on hybrid combination of CPU and GPU
capabilities

in a
computer environment in FAB01 through FAB10 computer
, which
installed with Nvid
ia
Quadro Fx 580 Graphics Cards
. As time perm
its,
both of
2
desired
symmetric and
asymmetric
architectures

approaches

should be explored
. Then use the run time
data to
determine

the

final

architecture

for the project
.




We have two basic approaches

(architectures)



Symmetric Approach



Divide the problem data, and perf
orm similar computations on
device and host



Asymmetric Approach



Put the problem data on the device



Have the device and host to different things (i.e., relaxation and
reduction)










The customer’s plan is to apply the rapid calculation performance an
d precision on
scien
tific research as a bench mark application. The standard is considered
rigorous and
challenging to achieve in the past.


After some discussion
s

with the

customer,

we collected a summary of the requirements.
We

also

prepare a project pla
n, schedule, and work estimate.


General Requirements


Basic features
:



Fully utilize the capabilities of the existing computer CPU and GPU hardware
environment to parallelize the computation through data parallel
.



Outperforms state of the art performance w
ith out
standing
c
alculation

efficiency
and m
aintain
s

at the top level

throughput

in ranking in the industry.



Visualize the calculation result in 3D graph format with GUI interface.



Scalable and
Maintainability
.


Options

features and efficient implementatio
ns
:



Use asynchronous transfers to com
municate interface node data at
each

iteration
so
as to
transfers overlap calculations.



Schedule the calculation for the interface nodes last so that the calculation can
progress during data transfer.



CUDA supports sim
ultaneous 2
-
way data transfer. Use it.



Use pinned memory for maximum data transfer speed (this may already be
enforced by CUDA asynchronous copy



After several iterations
, rebalance the CPU/GPU partitions based on observed
relative computation speed.


Work

products
:



Source code



Binary
code
for
computer
environment as
Fab01 to Fab10



Copies of internal work products



Cuda Profiler view of the run time data

such as bandwidthTest

and other
visualization



S
tart
GIT repository
ASAP

for the source code
, Documentat
ion, Schedule

and
enable the customer to a
ccess to repository.





Note that the customer

p
refer
s

Open Source License
.




Work Breakdown Structure (WBS)


For both of symmetric and asymmetric approaches, t
he work breakdown
structure
diagram shows the
Relaxati
on

project has been decomposed into
Device side and Host
side of implementations.

We partition the problem

set into device region and host region,
then communicate the device data to the GPU
.




Figure 1. WBS diagram








Schedule

and planning

The
proje
ct gets

started on July
13
, 2009 and will be finished on August 1
4, 2011
. The
first inception phase is from July
13

to
July 31
. The second round of
inception phase
starts on August 14 to add/explore

more

than one architectures option

to the project
.

Note
t
hat basic workflows diagram does not show the
time
over
lap

detail
for all phase
s but it
only shows the general

work sequence. The detailed schedule will show the project phase
overlap time information.



Figure 2.
Basic workflows









Figure3.
Detailed

schedule














Cost estimate
s

The project can be done in
158

human hours. Since it is a stude
nt project, the hourly cost
basis

is
estimated

as $50 US dollar/hour for a
n

inter
nship

rate.
Total cost of the
Jacobi
Relaxation

project is
79
00

US dolla
r.



Figure 4.
Labor hours and cost

breakdown