CIS 725 High Speed Networks. High-speed networks High bandwidth, high latency Low error rates Message loss mainly due to congestion.

CIS 725

High Speed Networks

High-speed networks

• High bandwidth, high latency• Low error rates• Message loss mainly due to congestion

• Example:

A and B connected via a 1 Gps link

RTD = 60ms

• Go Back N• Selective repeat

Parallelism

• Thread models

- thread per layer

- thread per module

- thread per message

- thread per connection

T3

T4

T1

T2

m1 m2 m3Thread per layer

- One thread for each layer- Pipelining

Thread per function

- One thread for each function- More than one function may be possible in a layer

Send function

Receive function

Thread per message

m1 m2

reordering

- Synchronization overhead

Thread pool

- Interface* recv_msg()

Thread per connection

• A separate thread is created for each connection• All messages for a connection are processed by the same thread

Blast Protocol

• Transmission occurs in blasts• Optimistic Assumption:

- Next packet to arrive is the expected one

- A blast is not interrupted by other packets

- Location of a data packet is known

DLL

APP

• Normal Path

- copy data packets into temp buffer

check header

copy into final location• Optimistic path

- copy data packet into final location

check header

if incorrect then do recovery

tempDLL

APP

header

data

• Selective retransmission based protocol

window = blast size

selective ack sent at end of blast

• If unexpected packet is received

- wait until end of blast

recopy items

send nack

Performance results

• Blast size = 32k• raw data rate = 9.4Mbs• Using normal processing = 63%• Best case optimistic = 85%• Worst case optimistic = 53%

Image Transfer Protocol

• Purpose = transfer of stored images• Many images itself have certain degree of

uncertainty• Most images have spatial redundancy• Neighboring pixels are highly correlated

• No retransmissions

- a message should not contain any two neighboring pixels

• Protocol must handle packets out of order and in parallel

• The source image is decomposed into messages, each containing info to deduce the position of the pixels

• Image is pre-processed and stored as exact messages to be transferred

• Congestion control:

- increases the transfer rate until the maximum acceptable loss rate is achieved