1. Reddy Mainampati Udit Parikh Alex Kardomateas

2. Introduction
How does application behavior impact wireless performance?
What can be done to improve performance?
A3: Application-Aware Acceleration for Wireless Data Networks
Application-aware but application transparent
Changes required by entities at each end

3. Evaluation Applications FTP – file transfer protocol
CIFS – printer and file sharing
SMTP – communication
HTTP – web pages
Traffic
IxChariot – software to generate traffic

4. Environment Wireless Networks WLAN – wireless local area
WWAN – wireless wide area
SAT – satellite area
Transport Protocols
TCP-ELN – NewReno + explicit loss notification
WTCP – Wide area wireless TCP
STP – Satellite transport protocol
Table of Parameters
Avg RTT
Avg Loss
Avg BW
WLAN
5 ms
1 %
5 Mbps
WWAN
200 ms
8 %
0.1 Mbps
SAT
1000 ms
3 %
1 Mbps

5. Analysis of Transport Protocols
FTP
CIFS

6. Analysis of Transport Protocols
SMTP
HTTP

7. Impact of Application Behavior
Thin session control messages
When thin message loss occurs, it takes one RTT to recover from the loss
When there is no packet following a lost thin message, loss may only be detected by RTO expiration
Batched data fetches
When requested data is smaller than BDP, there is underutilization of the network
Inflation of transaction time causes lower performance

8. Impact of Application Behavior
Flow control bottlenecked operations
If application reads slowly or freezes, receiver buffer can fill up causing communication to halt
Other reasons
Non-prioritization of data
Example: web browsing on a PDA
Non-use of data reduction techniques
Reducing content transferred between SMTP servers

9. A3 Design Why not change the applications to address these problems?
Rewrite applications or implement A3
A3 is an optimization module that is application-aware but application transparent
Transaction Prediction
Redundant and Aggressive Retransmission
Prioritized Fetching
Infinite Buffering
Application-aware Encoding

10. Transaction Prediction
Transaction Prediction (TP)
An approach to deterministically predict future application data requests to the server and issue them ahead of time
As the application requests new blocks, TP requests future blocks as well instead of waiting for the application
Blocks are cached by TP and sent when the application requests them
TP will aid in conditions where BDP is larger than the application batch fetch size and where RTT is large
For CIFS, TP requests the entire file at once instead of pieces
Benefits CIFS and HTTP

11. Redundant and Aggressive Retransmissions
Redundant and Aggressive Retransmissions (RAR)
An approach to protect thin messages from losses
RAR recognizes thin messages and uses a combination of packet level redundancy and aggressive retransmission to protect them
When A3 receives a message from the application, it checks to see if it is a thin message
If the message is a thin message, redundant copies are created, information about current time is stored, and retransmission timer is started
When a response arrives, the timestamp is checked and RTT is estimated from timestamp and message is forwarded to application
If retransmission timer expires for a thin message, the message is copied and retransmitted
A3 uses a timeout value of RTTavg + α
Benefits CIFS, SMTP, and HTTP

12. Prioritized Fetching Prioritized Fetching
An approach to prioritize data and transfer high priority data faster
Uses a separate module called application state monitor (ASM) which gets information from application
ASM classifies objects as high or low priority
ASM informs A3 about which objects are high priority and which are low priority
Data that is more important uses TCP-like congestion control for high priority information
Low priority congestion control (Ex. TCP-LP) for low priority data
Benefits HTTP and SMTP

13. Infinite Buffering Infinite Buffering (IB)
An approach to fix the flow control bottleneck by removing the limit on the receiver buffer
Uses secondary storage such as hard disk for buffer
IB keeps track of ACKs received by active TCP connection
If advertised window is less than maximum possible window size, IB changes the window size to max
If receivers buffer is full, IB stores data packets to disk cache and sends a proxy ACK back to the server
Once buffer has space, packets are sent from Disk cache to the buffer
When applications generates ACK for data that was already ACKed by IB, IB suppresses it
Benefits CIFS, SMTP, HTTP

14. Application-Aware Encoding
Application-Aware Encoding (AE)
An approach that uses application specific information to encode or compress data
Data compressed messages are marked so that the end user recognizes and de-compress data
Example: Table of words commonly used in s used to compress data when transmitting s
In A3 application vocabulary table are created on a user pair-wise basis (Ex. A-B vocabulary table created based on all past transmissions from A to B)
Benefits SMTP and HTTP

15. A3 Point Solution – A3● A3 is an end-to-end approach
A3● is a point solution and doesn’t require changes at both ends
AE cannot be used in A3● because compression and encoding requires an end-to-end approach
TP functions by generating more data requests and is effective as long as the server supports multiple simultaneous requests
RAR generates redundant transmissions and functions properly as long as the server can filter duplicate messages
If not, the redundant transmissions must be done in the transport layer, since TCP filters redundant messages

16. Performance Evaluation
CIFS
SMTP
HTTP
Integrated results for performance in WWAN

17. Related Work Odyssey Project
System support for collaboration between operating system and individual applications by letting them be aware of wireless environment
Coda file system
Designed to support disconnect operations for mobile hosts
Caches files at the client for disconnect periods

18. Critique
In performance evaluation, the integrated results were only tested in the WWAN environment
For the integrated results, A3 was modified for each protocol to support only three features instead of having A3 with every feature for each protocol which is more realistic
There were no performance evaluations done for A3●

19. Questions?