Loading...
Please wait, while we are loading the content...
Similar Documents
Handling High Bandwidth Aggregates By Receiver Driven Feedback Control ( Extended Abstract )
| Content Provider | Semantic Scholar |
|---|---|
| Author | Tan, Chee-Wei Chiu, Dah-Ming Lui, John C. S. Yau, David K. Y. |
| Abstract | Network congestion control solutions fall into one of two categories: end-to-end without involving intermediate routers, and hop-by-hop to shorten the control loop but at the expense of more costly router design and operation. The Internet protocol suite requires intelligence mainly at the network end points. TCP, for example, performs end-to-end control, where senders adjust their allowed windows of outstanding packets according to acks of previous data sent and receive windows advertised by receivers. Since the absence of a timely ack can be the result of either a congestion event or a corrupted packet, such windowbased congestion control may be unduly affected by the goal of error control [2]. To decouple congestion control from error control, a rate-based design (e.g., ATM ABR) typically feeds back a binary congestion control to multiple senders, who then adjust their sending rates accordingly. In this work, we consider a variant congestion control problem in which a receiver computes an explicit allowed traffic rate and uses it to regulate multiple routers sending to the receiver. Our design is thus rate-based, but uses an explicit rate signal instead of a binary signal. Such congestion control has many applications in practice, such as a web server regulating its Internet clients in a flash crowd scenario, or a server trying to protect itself from aggressive sources during a distributed denial-of-service (DDoS) attack. Intelligence is mostly required at the receiver since all the essential control decisions are made there. A subset of the routers participate in the control but do not make control decisions. As sources may not be trusted (e.g., in a DDoS attack), their participation is neither required nor assumed. In our approach, a network receiver, say S, experiencing resource overload installs a router throttle [6] at a set of upstream routers that are several hops away. The throttle specifies the maximum rate (in bits/s) at which packets destined for S can be forwarded by each router. Traffic that exceeds the rate limit will be dropped by the router. An installed throttle will generally change the load experienced at S. S can then adjust the throttle rate in multiple rounds of feedback control until it achieves its load target. Our goal is to derive a throttle algorithm that is (i) highly adaptive by avoiding unnecessary control parameters that would require configuration, (ii) able to converge quickly to a fair resource allocation, (iii) highly robust against extrinsic factors beyond the system's control (e.g., dynamic input traffic, and number/locations of current sources), and (iv) stable under given delay bounds. |
| File Format | PDF HTM / HTML |
| Alternate Webpage(s) | http://www.cs.cuhk.hk/~cslui/PUBLICATION/cybernetics2005.pdf |
| Alternate Webpage(s) | http://www.cs.cuhk.edu.hk/~cslui/PUBLICATION/cybernetics2005.pdf |
| Alternate Webpage(s) | http://www.cse.cuhk.edu.hk/~cslui/PUBLICATION/cybernetics2005.pdf |
| Language | English |
| Access Restriction | Open |
| Content Type | Text |
| Resource Type | Article |
