E. Cohen, H. Kaplan, and J. D. Oldham. Managing TCP connections under persistent HTTP. Computer Networks, 31:1709--1723, 1999.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....To test transparent proxies under a similar architecture. The current version requires explicit configuration to use a parent proxy. To consider the e#ects of significantly di#ering connection timeouts and replacement policies, as connection management policies will a#ect proxy performance [15]. To add management of DNS lookups. In the current implementation the first proxy doing the lookup will pay the penalty for a DNS miss, but not subsequent proxies, thus penalizing prefetching proxies inappropriately. Since DNS resolution costs can be severe [14] this will be an important ....

E. Cohen, H. Kaplan, and J. D. Oldham. Managing TCP connections under persistent HTTP. Computer Networks, 31:1709--1723, 1999.

.... along with a proliferation of commercial products [28, 29, 30, 34, 35] has mostly focused on the following areas: Web request distribution based on server load, content locality, cache affinity, or client session affinity [1, 14, 26, 28] Optimizations of server TCP connection management [4, 24, 34]. Mechanisms for inter connecting client and server TCP connections (TCP splicing) 3, 12, 19] Offloading server CPU intensive operations, such as secure sockets layer (SSL) operations. Each of these approaches has been based on optimizing and distributing web transactions by transparently ....

Edith Cohen et al. Managing TCP Connections under Persistent HTTP , Proc. of the Eighth International World Wide Web Conf., 1999.

....are using or are planning to use parallel TCP connections to increase aggregate TCP throughput for their application. The ubiquitous example of this is the Netscape Web browser, which uses an empirically determined value of four for the number of parallel TCP connections use by the client [25]. The GridFTP project [26] allows the user to select the number of parallel TCP connections to use for FTP data transfer. Storage Resource Broker (SRB) 27] has provisions to use multiple TCP sockets to improve SRB data transfer throughput. The Internet 2 Distributed Storage Initiative (I2DSI) ....

E. Cohen, H. Kaplan and J. Oldham, "Managing TCP Connections under Persistent HTTP", Proceedings of the Eighth International World Wide Web Conference, Toronto, Canada, May 1999

....timeout of 15 seconds, our approach reduces the maximum number of simultaneous connections by a third for the 4KB threshold and by 20 for the 1460 bytes threshold. The difference can be reduced by decreasing the timeout value, which has in fact been recommended for servers under high load [9] [31]. But that increases dramatically the total number of connections (Figure 4b) meaning that more requests must pay the overhead of connection establishment and slow start. Thus, DHTTP has a significant advantage over HTTP over the entire range of timeout values, with most benefits due to the ....

....5 Mean Latency (seconds) Number of Client Machines DHTTP(3 conn) DHTTP(1 conn) Apache(1 conn) Apache(3 conn) Altruistic Clients Modified Apache Fig. 10. Comparison of Apache and DHTTP servers: throughput (left) and latency (right) and request history indirectly address the same issue [31]. Further, SURGE clients never close connections, hoping to reuse them for future requests. To test a more altruistic client behavior, we modified SURGE to follow a recently proposed early close policy [9] where clients close their connections after getting all the files in one HTML page, thereby ....

E. Cohen, H. Kaplan, and J. D. Oldham, "Managing TCP connections under persistent HTTP," in Proceedings of the 8th International WWW Conference, (Toronto, Canada), May 1999.

....timeout of 15 seconds, our approach reduces the maximum number of simultaneous connections by a third for the 4KB threshold and by 20 for the 1460 bytes threshold. The difference can be reduced by decreasing the timeout value, which has in fact been recommended for servers under high load [7] [12]. But that increases dramatically the total number of connections (Figure 4b) meaning that more requests must pay the overhead of connection establishment and slow start. Thus, DHTTP has a significant advantage over HTTP over the entire range of timeout values, with most benefits due to the ....

....connections can be served. The resulting behavior is that the server uses high timeout when it has spare connections and switches to a sub second timeout otherwise. Recent proposals for dynamic adjustments of the timeout value based on the load and request history indirectly address the same issue [12]. Further, SURGE clients never close connections, hoping to 9 0 10000 20000 30000 40000 50000 60000 1 2 3 4 5 Number of Files Served Number of Client Machines DHTTP (3 conn) DHTTP (1 conn) Apache (1 conn) Apache (3 conn) Altruistic Clients Modified Apache 0 1 2 3 4 5 6 1 ....

Edith Cohen, Haim Kaplan, and Jeffrey D. Oldham. Managing TCP connections under persistent HTTP. In Proceedings of the 8th International WWW Conference, Toronto, Canada, May 1999.

....policies for Web servers may also be useful. For example, a Web server could automatically adjust the idle timeout value in order to keep the number of active sessions within a specified range. A number of TCP connection management policies for persistent HTTP have been examined by Cohen et al. [9]. Our previous work on Web server workload characterization [3] analyzed access logs that predated the widespread use of browsers with persistent (i.e. disk) caches, proxy caches, Performance Implications User Session Analyses Arlitt and Jin Page 83 of 90 and (transparent) network caches. Today, ....

E. Cohen, H. Kaplan and J. Oldham, "Managing TCP Connections under Persistent HTTP", Proceedings of the Eighth International World Wide Web Conference, Toronto, Canada, May 1999.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC