masters of science presentation: bringing the grid home

Post on 20-Nov-2014

1.411 Views

Category:

Technology

1 Downloads

Preview:

Click to see full reader

DESCRIPTION

Masters of Science presentation of my work on G-ICING

TRANSCRIPT

1

Bringing the Grid Home Master’s Thesis Presentation for Chris Sosa

University of VirginiaApril 28, 2009

Overview Motivation G-ICING Design Prototype Evaluation Demo Conclusion

2

Motivating Example: Biomedical Researcher

3

More Motivating Examples Many examples

Medical clinicians want patient records that complete and up-to-date

Researchers wants access to data provided at other institutions

Industry wants access to integrated customer and supply management data

Commonalities Lots of data to integrate – data is stronger together

than separate Store in various locations, with different access

control and security policies

4

Current Solutions

5

OR

A Better Solution: Data Grids Grid computing is a form of

distributed computing with Loosely coupled machines Machines cover multiple organizations

A Data Grid a type of Grid Computing system that deals with controlled sharing and management of large loads of data

6

7

Why don’t more people use Data Grids?

Hard to Use

Inflexible Security

Doesn’t play well with others

8

Solution CriteriaSimple

and Familiar

Flexible Security

Standards-Based

Perform well

enough

Simple and Familiar: More difficult than it seems

Often overlooked or treated as a secondary goal

Two aspects User Transparency Application

Transparency Solutions

Shell Extensions Shells Special Libraries Filesystems

9

10

Related Work OpenAFS creates a modified Samba server but stuck to

Samba/CIFS security model LUFS and FUSE are filesystem in user-space technologies

for UNIX / Mac Lack support for Windows Tied to UNIX security semantics

Gfarm uses FUSE + syscall hook library Same problems with just FUSE Overly complex for Windows, requires set up of a

separate Linux box to forward messages through Glite provides POSIX-like interface that is neither user or

application transparent

11

Bring in G-ICING Real filesystem for

Windows User transparency Application

transparency Full filesystem stack

so not tied to Windows security model

12

G-ICING Design

User Forwarding

Service (UFS)

Kernel Management

Service (KMS)

Inverted Calls

Grid-backend

Java VM

Grid Interface Service (GIS)

RN

S /

Byt

eIO

I/O Request JNI

13

G-ICING Design

UFSKMS Inverted Calls

Grid-backend

Java VM

GIS

RN

S /

Byt

eIO

I/O Request JNI

14

IFS Development in Windows

RDBSS (optional)I/O Sub SystemI/O RequestProgram

or OS I/O Request IFSDriver

15

G-ICING Design

UFSKMS Inverted Calls

Grid-backend

Java VM

GIS

RN

S /

Byt

eIO

I/O Request JNI

16

Kernel Management Service (KMS) Installable File System Driver

Network Redirector Kernel driver that interacts with other

Kernel components Communicates to User-mode UFS

with Inverted Call Model

17

User to Kernel CommunicationStep 1 Step 2

Step 3

UFS

KMS

Special

User Process

I/O

JNI

Grid-backend

Net

wor

kStep 4

Spe

cial

O/I

User Process

UFS GIS

KMS

KMS

KMS

18

G-ICING Design

UFSKMS Inverted Calls

Grid-backend

Java VM

GIS

RN

S /

Byt

eIO

I/O Request JNI

19

User Forwarding Service (UFS) Uses JNI to

communicate and forwards requests to GIS

Prompts user for credentials and obtains a delegated credential for use Flexible Security

through Delegation

20

G-ICING Design

UFSKMS Inverted Calls

Grid-backend

Java VM

GIS

RN

S /

Byt

eIO

I/O Request JNI

21

Grid Interface Service (GIS) Converts FS requests into ByteIO/RNS calls

Resource Naming Service (RNS) Basic directory services

ByteIO - files Caches meta-information ByteIO buffering In Java

Easy xml serialization/deserialization Problems with garbage collection

22

Prototype Implementation Genesis II as Grid-backend

Open-source Standards-based Developed at UVA

Semantics Time-out cache semantics – 45 seconds Write-through cache

23

Evaluation Performance: Do we perform well

enough? Usability

Compare to alternatives Usability Study

24

Performance: Evaluation Setup Client

Single-core 2.34 GHz desktop machine with 1GB memory running WinXP

100 Mbps connection Grid-Backend

Genesis II running on seven 8-core Xeon processors running at 2.33 GHz with 16 GB memory

1 Gbps connection

25

Performance: Test Plan Performance tests using Iozone Compare against Samba Share

Samba commonly used in organizations with shared filesystems

Compare G-ICING’s Iozone results with Samba results

26

(Re-)Write Performance G-ICING vs Samba

0

2000

4000

6000

8000

10000

12000

14000

64 128 256 512 1024 2048 4096 8192 16384 32768 65536

File Size (KB)

Thro

ughp

ut (K

B/s

)

Write Re-Write S_Write S_ReWrite

27

(Re-)Read G-ICING vs Samba

0

2000

4000

6000

8000

10000

12000

14000

16000

18000

20000

1024 2048 4096 8192 16384 32768 65536

File Size (KB)

Thro

ughp

ut (K

B/s)

Read Re-Read S_Read S_Re-Read

28

Usability Evaluation Alternatives?

Shell Extension (not app transparent) Posix-like libraries (neither user or app

transparent) Shell-like interfaces (not user

transparent) Web Portals (not app transparent)

Usability Study follows

29

Usability Study - Overview The Usability Study: Is the filesystem

paradigm really simple and familiar? 10 participants

6 non-engineering students in their first/second 3 graduate students with shell experience 1 user with knowledge of the Genesis II

30

Usability Study – A Look Inside Two tests run

Edit a MS Word document Run a “Grid-job” by copying a job description file

(JSDL) appropriately Each run either using G-ICING or the Genesis II

shell Questionnaire

Background How long each task took Measure how difficult each method was Give an overall preference

31

Usability Study - Results 9/10 users preferred G-ICING either

strongly or moderately Sole user who did not was also previous

Genesis II user Concerned with performance

6/9 felt strongly for G-ICING

32

Usability Study - Quantitative Results Average difficulty levels on a scale of 1-5

For Genesis II Shell – 3.6 For G-ICING – 1.3

Results below for duration of tasks

Shell Edit

Shell Run

G-ICING Edit

G-ICING Run

Overall: Avg. Duration (mins)

10 5.889 1.6 1.9

Shell Users: Avg. Duration (mins)

8 3 1.4 1.8

33

Demo

34

Conclusions The Grid is useless without users By providing a simple and familiar interface, G-ICING has

the potential to bring more users 9/10 prefer G-ICING over shell interface Takes five to ten times less time to perform common

data operations Provides both user and application transparency without

sacrificing flexible security, usage of standards and performance

35

Future Work1. Stretching filesystem paradigm to perform

more Grid services2. FUSE for Windows

36

Questions

?

37

Usability Study - Observations Graduate Students and Grid user had

higher expectations of Word Performance issues not discussed in

undergraduate session Most common complaint, “help” not

helpful

38

Prototype in Action

39

Prototype in Action (Continued)

top related