The development in Network PerformanceThe development in Network Performance

And it’s impact on the computing model of tomorrowAnd it’s impact on the computing model of tomorrow

The GRiD

• Named after the power-grid• Sometimes referred to as the

information power grid• Like the power-grid GRID should be

powered by large installations– not individual generators

Philosophy

• Current ’Internet’ only allows access to information

• The GRiD should provide access to any desired resource– CPU/SuperComputers– Storage– Applications

Performance Improvement since 1988

Net Latency

Net Bandwidth

Disk Latency

Disk Bandwidth

0

1

2

3

4

5

6

Imp

rov

em

en

t (o

rde

rs o

f m

ag

nit

ud

e)

Rules of the GameCopenhagen-Stockholm

• 1988– Latency 40 ms– Bandwidth 64 kb/s

• 2005– Latency 10 ms– Bandwidth 10Gb/s

• Networking is much better

Is There an Improvement?

Whether we have an improvement depend on our watch!

CPU

Whether we have an improvement depend on our watch!

Is There an Improvement?

CPU

OuterClock

InnerClock

Whether we have an improvement depend on our watch!

Is There an Improvement?

Rules of the GameCopenhagen-Stockholm

• Using the inner clock• 1988

– 1B: 0.8M CPU cycles– 1GB: 2T CPU cycles

• 2005– 1B: 39M CPU cycles– 1GB: 3G CPU cycles

• Latency is much worse• But bandwidth is much better

Rules of the GameHarddrive-Memory

• Using the inner clock• 1988

– 1B: 1M CPU cycles– 1GB: 1G CPU cycles

• 2005– 1B: 13M CPU cycles– 1GB: 38G CPU cycles

• Hard-drives are also much worse

Development as seen from the CPU

Net Latency

Net Bandwidth

Disk Latency

Disk Bandwidth

-2

-1.5

-1

-0.5

0

0.5

1

1.5

2

2.5

3

3.5

Imp

rov

em

en

t -

ord

ers

of

ma

gn

itu

de

Why is GRID?

• Network bandwidth is now here

Bandwidth Development

0.001

0.01

0.1

1

10

100

1000

10000

1974 1980 1988 1992 1995 1997 2000 2003

Year

Mb/

s LANWANDisk

Transparent Remote File Access

• Huge input files incur a number of problems:– Download time vs. total execution time– Job execution on the resource is delayed– Storage requirements on resources

• Often only small scattered fragments of input files are needed

• How about automatic on-demand download of needed data?

int fd = open(“inputfile”, O_RDONLY);while ((i=read(fd, &buffer, 2000)) >0){ /* process buffer */}

Example

User applications need not be recompiled or rewritten!

Communication Protocol

• HTTP supports a “range” parameter in get request:

GET /inputfile HTTP/1.1HOST:

MiG_server.imada.sdu.dkRange: bytes=2000-3000

• No range support in put requests– In order to support writing to remote files,

a custom web server is developed

Overriding file-access

• Override a subset of file manipulating routines:– open, close, read, write, seek, dup, sync, etc.

• Preload this library using the LD_PRELOAD environment variable– Requires user apps to be dynamically

linked

• Forward local file access to the native file system using the dlfcn library

Efficient Access

• Simple solution: general purpose block size based on n/2-analysis

• Advanced solution: depends on the user application:– The nature of the application (sequential vs non-

sequential file access)– The block size used in the application

• Introduce prefetching (1 block read-ahead)• Adjust the block size dynamically based on

the prefetching and the time taken to transfer a block

Experiments

• 4 experiments:– Overhead: read a one byte file– I/O intensive application: Checksum a 1 GB file– I/O balanced application: Process a 1 GB file– Partial file traversal: Search a 360 MB B+ tree for

a random key

• 3 test setups:– Local execution– Copy model– Remote access model

Baseline Performance100Mb net

Experiment Local Copy Remote

1B file 0.002 0.152 0.008

Checksum 50.11 130.1 114.3

Balanced 632.83 721.22 600.72

B+ Tree 0.002 30.692 0.0186

Latency tests

LatencyOdense 0 msCopenhagen 4 msÅrhus 9 msAalborg 13 msTromsø 32 msCanterbury 54 msChania 88 ms

Checksum

Balanced

B+ Tree

True End of the PC?

• If we can eliminate the disk we eliminate >60% of the errors in the PC

• But perhaps we don’t need the PC– The average PC utilizes less that 5% of

its capacity (Source: Intel)

• Reality is that the PC is– Much too powerful most of the time– Not nearly powerful enough the rest of

the time

• So we eliminate the PC?

Bandwidth for Remote users

• A graphics intensive user– Screen size: 1600x1400– Frequency: 50Hz– Color depth: 32b– Compression 1:10

• Required bandwidth: 0.33 Gb/s• Translates into 30 users per 10Gb line

Bandwidth for Remote users

• A typical user– Screen size: 1280x1024– Frequency: 30Hz– Color depth: 24b– Compression 1:100

• Required bandwidth: 0.008 Gb/s• Translates into 1138 users per 10Gb

line

World of Tomorrow?

GRID

User

Resource

Resource

GRIDDisk

The Grid Terminal

Grid terminal

But we have seen this before?

• Is this not just another thin client?• No!

– Thin clients work against dedicated servers

– Grid has no single point of failure– And Grid has competition

Distributed Shared Memory

DSM Test – the problem…

Points Latency (us) Bandwidth (MB/s)

SDU-SDU 130 63.0

NBI-NBI 250 95.3

SDU-NBI 1755 76.3

SDU-DIKU 1767 64.7

NBI-DIKU 3422 58.1

The Results

Conclusion and Predictions

• No reason to expect any change in the development of performance

• Networks will be increasingly slower• But bandwidth is limited only by

demand• Grid will allow users to ignore

computer maintenance and backups• Even individual home-users will join

Grid