[PPT] - Computing While Charging Building a Distributed Computing PowerPoint Presentation

SLIDE 1

Computing While Charging

Building a Distributed Computing Infrastructure Using Smartphones

Mustafa Y. Arslan - Indrajeet Singh - Shailendra Singh Harsha V. Madhyastha - Karthikeyan Sundaresan - Srikanth V. Krishnamurthy

Thursday, December 13, 12

SLIDE 2

Smartphones and Computing

Smartphones with higher CPU clock speeds, more

CPU cores, and so on.

real computers in our pockets
Enterprises are also adopting smartphones.
Problem: The real computing power of

smartphones is yet to be tapped into.

battery drains quickly -- long idle charging times

(e.g., at night)

Thursday, December 13, 12

SLIDE 3

Idle Phones Put Back to Work

Utilize idle charging periods for executing distributed jobs
n phones, for an enterprise.
100s-1000s of smartphones, working in parallel

Thursday, December 13, 12

SLIDE 4

Idle Phones Put Back to Work

Utilize idle charging periods for executing distributed jobs
n phones, for an enterprise.
100s-1000s of smartphones, working in parallel
Why would I bother? I have a datacenter full of PCs.

Thursday, December 13, 12

SLIDE 5

Idle Phones Put Back to Work

Utilize idle charging periods for executing distributed jobs
n phones, for an enterprise.
100s-1000s of smartphones, working in parallel
Why would I bother? I have a datacenter full of PCs.
untapped horsepower -- phones will run the same code

as your servers

Thursday, December 13, 12

SLIDE 6

Idle Phones Put Back to Work

Utilize idle charging periods for executing distributed jobs
n phones, for an enterprise.
100s-1000s of smartphones, working in parallel
Why would I bother? I have a datacenter full of PCs.
untapped horsepower -- phones will run the same code

as your servers

no cost to bootstrap (phones are already in hand)

Thursday, December 13, 12

SLIDE 7

Idle Phones Put Back to Work

Utilize idle charging periods for executing distributed jobs
n phones, for an enterprise.
100s-1000s of smartphones, working in parallel
Why would I bother? I have a datacenter full of PCs.
untapped horsepower -- phones will run the same code

as your servers

no cost to bootstrap (phones are already in hand)
a phone consumes ~5% of the energy of a PC.

Thursday, December 13, 12

SLIDE 8

Idle Phones Put Back to Work

Utilize idle charging periods for executing distributed jobs
n phones, for an enterprise.
100s-1000s of smartphones, working in parallel
Why would I bother? I have a datacenter full of PCs.
untapped horsepower -- phones will run the same code

as your servers

no cost to bootstrap (phones are already in hand)
a phone consumes ~5% of the energy of a PC.
no wiring, switching or cooling needed.

Thursday, December 13, 12

SLIDE 9

Idle Phones Put Back to Work

Utilize idle charging periods for executing distributed jobs
n phones, for an enterprise.
100s-1000s of smartphones, working in parallel
Why would I bother? I have a datacenter full of PCs.
untapped horsepower -- phones will run the same code

as your servers

no cost to bootstrap (phones are already in hand)
a phone consumes ~5% of the energy of a PC.
no wiring, switching or cooling needed.
Case to consider smartphones as a supplement for existing

computational systems.

Thursday, December 13, 12

SLIDE 10

job binary job binary

The Desired System

job binary

input

central server

Thursday, December 13, 12

SLIDE 11

job binary job binary

The Desired System

job binary

input

central server

Thursday, December 13, 12

SLIDE 12

job binary job binary

The Desired System

job binary central server

Thursday, December 13, 12

SLIDE 13

job binary job binary

The Desired System

job binary central server

X unplugged

Thursday, December 13, 12

SLIDE 14

job binary

The Desired System

job binary central server

X unplugged

Thursday, December 13, 12

SLIDE 15

The Desired System

central server

R E S U L T

Thursday, December 13, 12

SLIDE 16

The Desired System

central server

R E S U L T

Thursday, December 13, 12

SLIDE 17

Our Contribution

Thursday, December 13, 12

SLIDE 18

Our Contribution

Design & implement CWC.

Thursday, December 13, 12

SLIDE 19

Our Contribution

Design & implement CWC.
What is not novel: utilizing idle CPUs (e.g., Condor: A

Hunter of Idle Workstations [Litzkow, Livny, Mutka]).

Thursday, December 13, 12

SLIDE 20

Our Contribution

Design & implement CWC.
What is not novel: utilizing idle CPUs (e.g., Condor: A

Hunter of Idle Workstations [Litzkow, Livny, Mutka]).

What is novel: algorithm to optimally distribute computation

across smartphones with non-uniform bandwidths

Thursday, December 13, 12

SLIDE 21

Our Contribution

Design & implement CWC.
What is not novel: utilizing idle CPUs (e.g., Condor: A

Hunter of Idle Workstations [Litzkow, Livny, Mutka]).

What is novel: algorithm to optimally distribute computation

across smartphones with non-uniform bandwidths

Non-uniform wireless bandwidth calls for novel

schedulers (such as CWC).

Thursday, December 13, 12

SLIDE 22

Our Contribution

Design & implement CWC.
What is not novel: utilizing idle CPUs (e.g., Condor: A

Hunter of Idle Workstations [Litzkow, Livny, Mutka]).

What is novel: algorithm to optimally distribute computation

across smartphones with non-uniform bandwidths

Non-uniform wireless bandwidth calls for novel

schedulers (such as CWC).

Unique challenges not previously addressed.

100 10 20 30 40 Battery %

Time (Minutes)

Default Charging Curve Charge with CPU-intensive Jobs

Thursday, December 13, 12

SLIDE 23

Effect of Bandwidth on Scheduling

Files to be processed by the phones (queue a file

until next phone becomes available).

The server logs the service time (queueing +

transfer + processing) of each file.

server

Phones have identical CPUs but varying bandwidths

Queue of Files

Thursday, December 13, 12

SLIDE 24

Effect of Bandwidth on Scheduling

Files to be processed by the phones (queue a file

until next phone becomes available).

The server logs the service time (queueing +

transfer + processing) of each file.

server

Phones have identical CPUs but varying bandwidths

Queue of Files

Thursday, December 13, 12

SLIDE 25

Effect of Bandwidth on Scheduling

Files to be processed by the phones (queue a file

until next phone becomes available).

The server logs the service time (queueing +

transfer + processing) of each file.

server

Phones have identical CPUs but varying bandwidths

Queue of Files

Thursday, December 13, 12

SLIDE 26

Effect of Bandwidth on Scheduling

Two phones with the lowest bandwidths are

removed and the experiment is repeated.

server

Phones have identical CPUs but varying bandwidths

Queue of Files

Thursday, December 13, 12

SLIDE 27

0.5 1 1000 2000 3000 CDF Service Time (ms)

Too Much Parallelism Hurts

4 Phones 6 Phones

Thursday, December 13, 12

SLIDE 28

0.5 1 1000 2000 3000 CDF Service Time (ms)

Too Much Parallelism Hurts

Using only phones with high bandwidth can

compensate for reduced number of worker phones.

it is not a straight-forward choice to leverage the

full parallelism!

4 Phones 6 Phones

Thursday, December 13, 12

SLIDE 29

Problem Statement

Given a set of jobs and their input files, how do we

partition and distribute these files across a set of smartphones, to minimize the makespan?

Time

Local Execution Time (depends on CPU clock speed, the job itself & the input partition size) File Transfer Time (depends on bandwidth and the input partition size)

Thursday, December 13, 12

SLIDE 30

Problem Statement

Given a set of jobs and their input files, how do we

partition and distribute these files across a set of smartphones, to minimize the makespan?

Time

Local Execution Time (depends on CPU clock speed, the job itself & the input partition size) File Transfer Time (depends on bandwidth and the input partition size)

Thursday, December 13, 12

SLIDE 31

Problem Statement

Given a set of jobs and their input files, how do we

partition and distribute these files across a set of smartphones, to minimize the makespan?

Time

Local Execution Time (depends on CPU clock speed, the job itself & the input partition size) File Transfer Time (depends on bandwidth and the input partition size)

Thursday, December 13, 12

SLIDE 32

Problem Statement

Given a set of jobs and their input files, how do we

partition and distribute these files across a set of smartphones, to minimize the makespan?

Time

Local Execution Time (depends on CPU clock speed, the job itself & the input partition size) File Transfer Time (depends on bandwidth and the input partition size)

Thursday, December 13, 12

SLIDE 33

Problem Statement

Given a set of jobs and their input files, how do we

partition and distribute these files across a set of smartphones, to minimize the makespan?

Time

Local Execution Time (depends on CPU clock speed, the job itself & the input partition size) File Transfer Time (depends on bandwidth and the input partition size)

Makespan

Thursday, December 13, 12

SLIDE 34

Predicting Execution Times

Thursday, December 13, 12

SLIDE 35

Predicting Execution Times

bandwidth to each phone periodically measured. but

cannot measure the local execution time for every job- phone pair.

Thursday, December 13, 12

SLIDE 36

Predicting Execution Times

bandwidth to each phone periodically measured. but

cannot measure the local execution time for every job- phone pair.

run each job j using the slowest phone in the system:

e.g., with S MHz CPU

Thursday, December 13, 12

SLIDE 37

Predicting Execution Times

bandwidth to each phone periodically measured. but

cannot measure the local execution time for every job- phone pair.

run each job j using the slowest phone in the system:

e.g., with S MHz CPU

if the slowest phone takes Ts ms, then another phone

with A MHz CPU should take Ts * S / A ms.

Thursday, December 13, 12

SLIDE 38

Predicting Execution Times

bandwidth to each phone periodically measured. but

cannot measure the local execution time for every job- phone pair.

run each job j using the slowest phone in the system:

e.g., with S MHz CPU

if the slowest phone takes Ts ms, then another phone

with A MHz CPU should take Ts * S / A ms.

Thursday, December 13, 12

SLIDE 39

Predicting Execution Times

bandwidth to each phone periodically measured. but

cannot measure the local execution time for every job- phone pair.

run each job j using the slowest phone in the system:

e.g., with S MHz CPU

if the slowest phone takes Ts ms, then another phone

with A MHz CPU should take Ts * S / A ms.

Thursday, December 13, 12

SLIDE 40

Predicting Execution Times

bandwidth to each phone periodically measured. but

cannot measure the local execution time for every job- phone pair.

run each job j using the slowest phone in the system:

e.g., with S MHz CPU

if the slowest phone takes Ts ms, then another phone

with A MHz CPU should take Ts * S / A ms.

wrong estimates are corrected using execution reports sent to the central server.

Thursday, December 13, 12

SLIDE 41

Minimum Height Bin Packing

Given a finite set of items U, a size for each

item in U and a bin capacity C.

partition U into disjoint sets U1, U2, .., Un s.t.

the sum of the item sizes in each Ui is <= C.

Thursday, December 13, 12

SLIDE 42

Minimum Height Bin Packing

Given a finite set of items U, a size for each

item in U and a bin capacity C.

partition U into disjoint sets U1, U2, .., Un s.t.

the sum of the item sizes in each Ui is <= C.

Bin 1 Bin 2 Bin 3 C

Thursday, December 13, 12

SLIDE 43

Minimum Height Bin Packing

Given a finite set of items U, a size for each

item in U and a bin capacity C.

partition U into disjoint sets U1, U2, .., Un s.t.

the sum of the item sizes in each Ui is <= C.

Bin 1 Bin 2 Bin 3 C

Thursday, December 13, 12

SLIDE 44

Minimum Height Bin Packing

Given a finite set of items U, a size for each

item in U and a bin capacity C.

partition U into disjoint sets U1, U2, .., Un s.t.

the sum of the item sizes in each Ui is <= C.

Bin 1 Bin 2 Bin 3 C

Thursday, December 13, 12

SLIDE 45

Minimum Height Bin Packing

Bin 1 Bin 2 Bin 3

Thursday, December 13, 12

SLIDE 46

Minimum Height Bin Packing

Bin 1 Bin 2 Bin 3

Thursday, December 13, 12

SLIDE 47

Minimum Height Bin Packing

Each job input is an item (not rigid, i.e., can be partitioned & packed in different bins)

Bin 1 Bin 2 Bin 3

Thursday, December 13, 12

SLIDE 48

Minimum Height Bin Packing

Each job input is an item (not rigid, i.e., can be partitioned & packed in different bins)
Cost of partitioning (transferring files adds height depending on phone bandwidth)

Bin 1 Bin 2 Bin 3

Thursday, December 13, 12

SLIDE 49

Minimum Height Bin Packing

Each job input is an item (not rigid, i.e., can be partitioned & packed in different bins)
Cost of partitioning (transferring files adds height depending on phone bandwidth)
Phones are bins (but they are not identical)

Bin 1 Bin 2 Bin 3

Thursday, December 13, 12

SLIDE 50

Minimum Height Bin Packing

Each job input is an item (not rigid, i.e., can be partitioned & packed in different bins)
Cost of partitioning (transferring files adds height depending on phone bandwidth)
Phones are bins (but they are not identical)

Bin 1 Bin 2 Bin 3

Items occupy different heights depending on the bin they are packed in.
e.g., items behave like liquids.

Thursday, December 13, 12

SLIDE 51

Scheduling Jobs

C

Sorted List of Inputs

Thursday, December 13, 12

SLIDE 52

Scheduling Jobs

C

Sorted List of Inputs

Thursday, December 13, 12

SLIDE 53

Scheduling Jobs

C

Sorted List of Inputs

Thursday, December 13, 12

SLIDE 54

Scheduling Jobs

C

Sorted List of Inputs

Thursday, December 13, 12

SLIDE 55

Scheduling Jobs

C

Sorted List of Inputs

Thursday, December 13, 12

SLIDE 56

Scheduling Jobs

C

Sorted List of Inputs

Thursday, December 13, 12

SLIDE 57

Scheduling Jobs

C

Sorted List of Inputs

Thursday, December 13, 12

SLIDE 58

Scheduling Jobs

C

Sorted List of Inputs

Thursday, December 13, 12

SLIDE 59

Scheduling Jobs

C

Sorted List of Inputs

Thursday, December 13, 12

SLIDE 60

Scheduling Jobs

C

Sorted List of Inputs

Thursday, December 13, 12

SLIDE 61

Scheduling Jobs

try to produce few partitions to reduce the aggregation load at the central

server, while minimizing C.

C

Sorted List of Inputs

Thursday, December 13, 12

SLIDE 62

Automating Job Execution

We implement the CWC service in Android
runs in “background” -- no human input
exploits the compatibility between JVM and

Dalvik (a core subset of Java APIs are common)

leverages Java Reflection API to dynamically load

classes and execute methods defined by them

Thursday, December 13, 12

SLIDE 63

Automating Job Execution

We implement the CWC service in Android
runs in “background” -- no human input
exploits the compatibility between JVM and

Dalvik (a core subset of Java APIs are common)

leverages Java Reflection API to dynamically load

classes and execute methods defined by them

.java .class .dex load exec Server (Traditional Java) Phone (Android)

Thursday, December 13, 12

SLIDE 64

Automating Job Execution

We implement the CWC service in Android
runs in “background” -- no human input
exploits the compatibility between JVM and

Dalvik (a core subset of Java APIs are common)

leverages Java Reflection API to dynamically load

classes and execute methods defined by them

The same Java code runs on both PCs and

smartphones!

.java .class .dex load exec Server (Traditional Java) Phone (Android)

Thursday, December 13, 12

SLIDE 65

Setup

18 Android smartphones with CWC software.
Lightweight central server
Amazon EC2 small instance ( < 2 GB RAM )
Multi-threaded Java NIO implementation
Workload: 3 types of tasks
1. Count the prime numbers in a file (50 files)
2. Count the target word occurrences in a file (50 files)
3. Blur pixels in a photo (atomic -- 50 photos)

Connectivity 802.11a / g, EDGE, 3G, 4G CPU Speed 806 MHz to 1.5 GHz Single and Dual Core

Thursday, December 13, 12

SLIDE 66

Results

Shows a sub-set

f the phones

Thursday, December 13, 12

SLIDE 67

Results

Balanced assignment for phones 4, 12, 13, 14

(and for other phones not shown) Shows a sub-set

f the phones

Thursday, December 13, 12

SLIDE 68

Results

Balanced assignment for phones 4, 12, 13, 14

(and for other phones not shown)

Phones 2 and 9 finish earlier than others

(because they are “faster” than predicted). Shows a sub-set

f the phones

Thursday, December 13, 12

SLIDE 69

Results

Makespan is 1120 seconds.
%88 of the jobs are not partitioned (i.e.,

running on one phone), %9 have 3 partitions and %3 have 4 partitions.

Balanced assignment for phones 4, 12, 13, 14

(and for other phones not shown)

Phones 2 and 9 finish earlier than others

(because they are “faster” than predicted). Shows a sub-set

f the phones

Thursday, December 13, 12

SLIDE 70

Results

Makespan is 1120 seconds.
%88 of the jobs are not partitioned (i.e.,

running on one phone), %9 have 3 partitions and %3 have 4 partitions.

Balanced assignment for phones 4, 12, 13, 14

(and for other phones not shown)

Phones 2 and 9 finish earlier than others

(because they are “faster” than predicted). Shows a sub-set

f the phones
How about full parallelism?
each job has |P| partitions (one partition per

phone) -- makespan is 1720 seconds.

each job to a phone (as a whole) in a RR

fashion -- makespan is 1805 seconds.

Thursday, December 13, 12

SLIDE 71

THANK YOU!

QUESTIONS?

Thursday, December 13, 12