Henrik Frystyk Nielsen, James Gettys, Anselm Baird-Smith, Eric Prud'hommeaux, Hakon Wium Lie, and Chris Lilley. Network performance effects of HTTP/1.1, CSS1, and PNG. In Proceedings of ACM SIGCOMM'97 Conference, September 1997.  @ NUS  Home/Search   Document Not in Database   Summary   Related Articles   Check

....over the same TCP connection, amortizing the TCP set up overhead. Pipelining lets the client send multiple requests over the same connection without waiting for responses. The server will send a stream of responses back. These features have been shown to reduce client latency and network traffic [25]. However, they do not eliminate all overheads of TCP, and in fact may introduce new performance penalties, especially when the bottleneck is at the server [7] Persistent connections increase the number of open connections at the server, which can have a significant negative effect on server ....

Henrik Frystyk Nielsen, James Gettys, Anselm Baird-Smith, Eric Prud'hommeaux, Hakon Wium Lie, and Chris Lilley. Network performance effects of HTTP/1.1, CSS1, and PNG. In Proceedings of ACM SIGCOMM'97 Conference, September 1997.

....to these problems does not appear to have changed. We have informed the Apache developers of the problems and fixes we discuss below; we expect that some of our patches will be part of a future Apache release. Pipelining requests across a P HTTP connection is necessary to maximize performance [15]. Our simple client does not pipeline requests, but pipelining (by itself) would not eliminate any of the interactions we describe. Pipelining may reduce the effect of the shortinitial segment problem to one delayed ACK stall depending on the implementation. 2.2 The Short Initial Segment ....

Henrik Frystyk Nielsen, Jim Gettys, Anselm Baird-Smith, Eric Prud'hommeaux, Hakon Wium Lie, and Chris Lilley. Network performance effects of HTTP/1.1, CSS1, and PNG. NOTE-pipelining-970207, availble as web page http://www.w3.org/pub/- WWW/Protocols/HTTP/Performance/- Pipeline.html, 7 February 1997.

.... Points ffl General Web Performance Studies [14, 40] ffl Web Client Behavior Studies [9, 16, 20, 28, 32, 34, 52, 22] ffl Web Server Behavior Studies [3, 4, 5, 6, 7, 12, 13, 26, 29, 45] ffl Web Proxy and Caching Studies [2, 15, 23, 25, 36, 38, 51, 56, 57] ffl Network Effects of Web Traffic [11, 17, 19, 37, 44, 43, 42, 24, 27] ffl Load Generators [1, 8, 10, 18, 39, 46, 54, 55] # Web Server Performance Analysis Paul Barford Page 7 Key Questions for Performance Analysis ffl What performance metrics are important Functional correctness Sizing, tuning and capacity planning ffl How will the test ....

H. Frystyk-Nielsen, J. Gettys, A. Baird-Smith, E. Prud'hommeaux, H. Wium-Lie, and C. Lilley. Network performance effects of HTTP/1.1, CSS1 and PNG. In Proceedings of ACM SIGCOMM '97, Cannes, France, Setpember 1997.

.... indicate that connection establishment is a significant component of request response time [8] and that the short TCP connection associated with a single HTTP request rarely gets past the slow start phase [12] In order to improve user perceived latency, a case was made for persistent HTTP [18, 17]. A persistent TCP connection is kept open and re used to carry imminent future HTTP requests. The common older HTTP 1.0 supports a keep alive option which instructs the server to keep the TCP connection open beyond the handling of the current request. With the newer HTTP 1.1, persistence is a ....

Henrik Frystyk Nielsen, James Gettys, Anselm Baird-Smith, Eric Prud'hommeaux, Hakon Wium Lie, and Chris Lilley. Network performance effects of HTTP/1.1, CSS1, and PNG. In Proceedings of the ACM SIGCOMM'97 Conference, Cannes, France, August 1997.

....pages contain 20 to 40 embedded images. Following standard HTTP, the browser parses the containing HTML resource and sends separate GET request for each of the embedded resources. Most browsers reduce the total retrieval time by sending 4 to 5 requests in parallel and reusing the TCP connections [34]. However, even with such optimizations, downloading a typical Web page leads to at least 4 to 5 GET request rounds. The batch transmission method reduces the entire process into a single round with a large response. Batch transmission is implemented in our Web prototype and we are in the ....

....25(2) 1993. 32] S. Manley and M. Seltzer. Web facts and fantasy. In Proceeding of the USENIX Symposium on Internet Technologies and Systems, 1998. 33] M. Mitzenmacher. How useful is old information Technical Report TR 1998 002, Systems Research Center, Digital Equipment Corporation, 1998. [34] H. Frystyk Nielsen, J. Gettys, A. Baird Smith, Eric Prud hommeaux, H. Wium Lie, and C. Lilley. Network performance effects of HTTP 1.1, CSS1, and PNG. In Proceeding of ACM SIGCOMM 97, 1997. 35] M. Sayal, Y. Breitbart, P. Scheuermann, and R. Vingralek. Selection algorithms for replicated web ....

[Article contains additional citation context not shown here]

H. Frystyk Nielsen, J. Gettys, A. Baird-Smith, Eric Prud'hommeaux, H. Wium Lie, and C. Lilley. Network performance effects of HTTP/1.1, CSS1, and PNG. In Proceeding of ACM SIGCOMM'97, 1997.

....appears that nobody had attempted to quantify the potential benefits of loss free compression, although the GloMop project [7] did explore the use of lossy compression. A study done at the World Wide Web Consortium reports on the benefits of compression in HTTP, but for only one example document [15]. Also, our traces suggest that few existing client implementations offer to accept compressed encodings of arbitrary responses (apparently, Lynx is the one exception) Before the Web was an issue, Douglis [4] wrote generally about compression in distributed systems. The WebExpress project [9] ....

Henrik Frystyk Nielsen, James Gettys, Anselm Baird-Smith, Eric Prud'hommeaux, Hakon Wium Lie, and Chris Lilley. Network Performance Effects of HTTP/1.1, CSS1, and PNG. In Proc. SIGCOMM '97. Cannes, France, September, 1997.

....of these delays. The final problem is that the link level approach requires exact duplicate packet payloads, but the support for persistent connections and pipelining in HTTP 1. 1 provides a strong incentive for packing of multiple responses into one packet, and without regard for packet boundaries [38]. This may lead to unique packet payloads even with lots of HTTPlevel duplication, although it is also possible that HTML level page compositions will be static enough that substantial packet level duplication persists. Broder et al. 7] looked for syntactically similar documents in the Web, ....

Henrik Frystyk Nielsen, James Gettys, Anselm Baird-Smith, Eric Prud'hommeaux, Hakon Wium Lie, and Chris Lilley. Network Performance Effects of HTTP/1.1, CSS1, and PNG. In Proc. SIGCOMM '97. Cannes, France, September, 1997.

....pages contain 20 to 40 embedded images. Following standard HTTP, the browser parses the containing HTML resource and sends separate GET request for each of the embedded resources. Most browsers reduce the total retrieval time by sending 4 to 5 requests in parallel and reusing the TCP connections [34]. However, even with such optimizations, downloading a typical Web page leads to at least 4 to 5 GET request rounds. The batch transmission method reduces the entire process into a single round with a large response. Batch transmission is implemented in our Web prototype and we are in the ....

H. Frystyk Nielsen, J. Gettys, A. Baird-Smith, Eric Prud'hommeaux, H. Wium Lie, and C. Lilley. Network performance effects of HTTP/1.1, CSS1, and PNG. In Proceeding of ACM SIGCOMM'97, 1997.

....time. Web browsing sessions typically consist of many HTTP requests, each for a small size document. Practice with HTTP 1. 0 was to use a separate TCP connection for each HTTP request and response [7] Hence, incurring connectionestablishment and slow start 1 latencies on each request [17]. Persistent connections [23] address that by re using a single long lived TCP connection for multiple HTTP requests. Persistent connections became a default with HTTP 1.1 [16] which gets increasingly deployed. Deployment of HTTP 1.1 reduces the latency incurred in subsequent requests to a server ....

H. Frystyk Nielsen, J. Gettys, A. Baird-Smith, E. Prud'hommeaux, H. W. Lie, and C. Lilley. Network performance effects of HTTP/1.1, CSS1, and PNG. In Proceedings of the ACM SIGCOMM'97 Conference, Cannes, France, August 1997.

....time, HTTP request response time, server processing, and finally, transmission time. Practice with HTTP 1. 0 was to use a separate TCP connection for each HTTP request and response [6] Hence, incurring connection establishment and slow start [17] 1 latencies on each request [15]. Persistent connections [22] address that by re using a single long lived TCP connection for multiple HTTP requests. Persistent connections became a default with HTTP 1.1 [14] which gets increasingly deployed. Deployment of HTTP 1.1 reduces the latency incurred in subsequent requests to a server ....

H. Frystyk Nielsen, J. Gettys, A. Baird-Smith, E. Prud'hommeaux, H. W. Lie, and C. Lilley. Network performance effects of HTTP/1.1, CSS1, and PNG. In Proceedings of the ACM SIGCOMM'97 Conference, Cannes, France, August 1997.

....such as caching and distributed Web content. Our study is significant in that it examines the performance of various protocol options by measuring end to end response for actual Web servers from a variety of client sites. This is a significantly broader study compared to that of Nielsen, et al. [18], which measured the performance of HTTP 1.0 and HTTP 1.1 in a controlled setting for a single synthesized Web page. Our study also examines the contribution of other factors to end to end performance, such as caching, multi server content and byte range requests. We gathered a set of active ....

....be used for serialized request response pairs or for pipelining multiple requests and then receiving multiple responses. The relative performance of these protocol options was studied by Nielsen, but it was done on a limited basis with a single artificially created page with 40 embedded objects [18]. The experiment was useful to isolate specific low level issues on the impact of TCP on HTTP 1.1 and demonstrated the usefulness of pipelining persistent connections under the test conditions. A primary goal of our work is to study these protocol options using real life servers. ffl Caching. We ....

[Article contains additional citation context not shown here]

Henrik Frystyk Nielsen, Jim Gettys, Anselm Baird-Smith, Eric Prud'hommeaux, Hikon Lie, and Chris Lilley. Network performance effects of HTTP/1.1, CSS1, and PNG. In Proceedings of the ACM SIGCOMM '97 Conference. ACM, September 1997. http://www.acm.org/sigcomm/sigcomm97/papers/p102.html.

....and processing costs for piggybacked validations invalidations can be separated from the critical path cost of servicing the request. The persistent connection feature of HTTP 1. 1 is expected to have a major impact on the performance of the World Wide Web by reducing network overhead and latency [15]. The importance of cache replacement and coherency, even content caching itself [2] need to be considered in light of this new development. We have made a preliminary study on the performance of replacement and coherency policies in light of persistent connections. Using an approach similar to ....

Henrik Frystyk Nielsen, Jim Gettys, Anselm Baird-Smith, Eric Prud'hommeaux, Hikon Lie, and Chris Lilley. Network performance effects of HTTP/1.1, CSS1, and PNG. In Proceedings of the ACM SIGCOMM '97 Conference. ACM, September 1997. http://www.acm.org/sigcomm/sigcomm97/papers/p102.html.

....local loads on our server. Surge is a synthetic workload generator which is unlike previous workload generators because it incorporates a wide range or workload characteristics that are important to many different aspects of server performance. Studies of the network effects of Web traffic include [7, 8, 21, 30, 38, 43, 54, 56]. These studies show the performance effects of both HTTP protocol interaction and TCP packet level interactions. The shortfall of these studies when it comes to analyzing latency is that they only take measurements in one place (somewhere near an end point of an HTTP transaction) and, thus, ....

H. Frystyk-Nielsen, J. Gettys, A. Baird-Smith, E. Prud'hommeaux, H. Wium-Lie, and C. Lilley. Network performance effects of HTTP/1.1, CSS1 and PNG. In Proceedings of ACM SIGCOMM '97, Cannes, France, Setpember 1997.

....a minimum number of data packets. These enhancements, along with others, paved the way for the HTTP 1.1 specification which was proposed as an IETF standard in 1997 [12] Initial evaluation of HTTP 1. 1 indicated that under certain conditions it can reduce both client latency and network traffic [13]. However, the enhancements in HTTP 1.1 come at a cost. Because connections are persistent, servers must typically manage a much larger number of open connections. The work in [23] points out that requiring the server to manage many open connections could increase its processing and memory load. ....

....possible into each packet and placing separator strings between distinct files within a packet. Pipelining has the effect of reducing the number of request and response packets, thus reducing the number of network round trips required to complete a Web request response transaction. Nielsen et al. [13] show results indicating that pipelining is necessary in order to significantly reduce latency when using HTTP 1.1. A number of studies have analyzed the performance aspects of HTTP. In particular, 9] compares the performance of HTTP 1.0 versus HTTP 1.1 on a high latency wireless network and ....

[Article contains additional citation context not shown here]

H. Frystyk-Nielsen, J. Gettys, A. Baird-Smith, E. Prud'hommeaux, H. Wium-Lie, and C. Lilley. Network performance effects of HTTP/1.1, CSS1 and PNG. In Proceedings of ACM SIGCOMM '97, Cannes, France, September 1997.

.... but how does the use of embedded images change as these container resources change Does the set of embedded images change or is the set of images relatively constant with other aspects of the container page changing Do cascading style sheets, proposed to replace some uses of embedded images [19], exhibit similar change characteristics ffl Another area for study is to examine the predictability and locality of changes to a resource. This is particularly important for resources that change often such as dynamically computed content. Techniques such as delta encoding [17] and active ....

Henrik Frystyk Nielsen, Jim Gettys, Anselm Baird-Smith, Eric Prud'hommeaux, Hikon Lie, and Chris Lilley. Network performance effects of HTTP/1.1, CSS1, and PNG. In Proceedings of the ACM SIGCOMM '97 Conference. ACM, September 1997. http://www.acm.org/sigcomm/sigcomm97/papers/p102.html.

....introduced by the proxy a configurable cache miss time overhead and a configurable cache hit time overhead are added to the request service time. 3. 2 Network Connections As mentioned in the introduction, it is important to handle the interactions between HTTP and TCP, particularly persistent HTTP [7, 12, 15]. In PROXIM, each simulated client maintains zero or more open connections to the proxy. An idle connection is one in which no HTTP request is currently active. When a persistent HTTP request is generated and an idle connection exists for the client in question, an idle connection is chosen to ....

....even though the cache hit ratio was roughly 50 . However, their particular study focussed on corporate LANs and WANs, so they did not consider the particular highly disparate environment we have evaluated here. Other approaches to latency reduction include prefetching [17, 10] data compression [14, 15], and delta encoding [9, 1, 14] Duska, et al. 5] used a variety of proxy cache logs to evaluate hit rates, locality, sharing, and cache coherence. They found that proxy cache hit rates varied from 24 to 45 , with higher request rates resulting in higher hit rates. They focused on resource hit ....

[Article contains additional citation context not shown here]

Henrik Frystyk Nielsen, James Gettys, Anselm Baird-Smith, Eric Prud'hommeaux, Hakon Wium Lie, and Chris Lilley. Network performance effects of HTTP/1.1, CSS1, and PNG. In Proceedings of ACM SIGCOMM'97 Conference, pages 155--166, September 1997.

....As mentioned in section II B, HTTP 1.1 reduces internal fragmentation by filling up packets with more data. However, since the maximum packet size 1500 bytes for Ethernet is 1500 bytes and 1152 bytes for Ricochet, it needs to send more packets over Ricochet. V. Related Work Nielsen, et al. [5] describe the performance improvements of HTTP 1.1 over HTTP 1.0 on a variety of network technologies, but none of them are wireless. They do measure the performance of HTTP 1.1 on a high latency, low bandwidth network (a modem) but do not measure the performance of HTTP 1.0 for comparison. 7 ....

Henrik Frystyk Nielsen, Henrik Frystyk, Jim Gettys, Anselm Baird-Smith, Eric Prudhommeaux, Hakon Wium Lie, and Chris Lilley. Network performance effects of http/1.1, css1, and png. In Proceedings of SIGCOMM, 1997. http://www.w3.org/Protocols/HTTP/Performance/Pipeline.

....local loads on our server. Surge is a synthetic workload generator which is unlike previous workload generators because it incorporates a wide range or workload characteristics that are important to many different aspects of server performance. Studies of the network effects of Web traffic include [7, 8, 21, 29, 53, 37, 42, 52]. These studies show the performance effects of both HTTP protocol interaction and TCP packet level interactions. The shortfall of these studies when it comes to analyzing latency is that they only take measurements in one place (somewhere near an end point of an HTTP transaction) and, thus, ....

H. Frystyk-Nielsen, J. Gettys, A. Baird-Smith, E. Prud'hommeaux, H. Wium-Lie, and C. Lilley. Network performance effects of HTTP/1.1, CSS1 and PNG. In Proceedings of ACM SIGCOMM '97, Cannes, France, Setpember 1997.

....minimum number of data packets. These enhancements, along with others, paved the way for the HTTP 1.1 specification which was proposed as an IETF standard in 1997 [13] Initial evaluation of HTTP 1. 1 indicated that un der certain conditions it can reduce both client latency and network traffic [14]. However, the enhancements in HTTP 1.1 come at a cost. Because connections are persistent, servers must typically manage a much larger number of open connections. The work in [25] points out that requiring the server to manage many open connections could increase its processing and memory load. ....

....into each packet and placing separator strings between distinct files within a packet. Pipelining has the effect of reducing the number of request and response packets, thus reducing the number of network round trips required to complete a Web request response transaction. Nielsen et al. [14] show results indicating that pipelining is necessary in order to significantly reduce latency when using HTTP 1.1. A number of studies have analyzed the performance aspects of HTTP. In particular, 10] compares the performance of HTTP 1.0 versus HTTP 1.1 on a high latency wireless network and ....

[Article contains additional citation context not shown here]

H. Frystyk-Nielsen, J. Gettys, A. Baird-Smith, E. Prud'hommeaux, H. Wium-Lie, and C. Lilley. Network performance effects of HTTP/1.1, CSS1 and PNG. In Proceedings of ACM SIGCOMM '97, Cannes, France, September 1997.

....overhead and a configurable cache hit time overhead are added to the request service time (during cache hits and cache misses, respectively) 3. 2 Network Connections As we argue in the introduction, it is important to handle the interactions between HTTP and TCP, particularly persistent HTTP [3, 7, 9]. In PROXIM, each client maintains zero or more open connections to the proxy. An idle connection is one in which no request is currently active. When a persistent HTTP request is generated and an idle connection exists for the client in question, an idle Proxy Clients Servers modem queues ....

Henrik Frystyk Nielsen, James Gettys, Anselm Baird-Smith, Eric Prud'hommeaux, Hakon Wium Lie, and Chris Lilley. Network performance effects of HTTP/1.1, CSS1, and PNG. In Proceedings of ACM SIGCOMM'97 Conference, pages 155--166, September 1997.

....as possible. 5 Related Work Measuring and analyzing the performance of Web servers is an increasingly popular research topic. Existing research on improving Web performance has focused largely on reducing network latency, primarily through caching techniques [15, 27] or protocol optimizations [14, 19, 17]. In addition, the workload of Web servers have been modeled analytically at the University of Saskatchewan [12] SGI s WebSTONE is widely considered as the standard benchmarking system for measuring the performance of Web servers [6] and it is the basis of our own benchmarking methodology. A ....

....designs. Another way to improve Web performance is by removing overhead in the protocol itself. The W 3 C has recently standardized HTTP 1.1, which enables multiple requests over a single connection. This connection caching strategy can significantly enhance the performance over HTTP 1. 0 [25, 19, 17]. The need for persistent connections to improve latency was noted by Mogul [14] Latency can also be improved by using caching proxies and caching clients, although the removal policy needs to be carefully considered [27] JAWS can be extended to become a caching proxy, allowing it to leverage ....

Henrik Frystyk Nielsen, Jim Gettys, Anselm Baird-Smith, Eric Prud'hommeaux, Hakon Wium Lie, and Chris Lilley. Network Performance Effects of HTTP/1.1, CSS1, and PNG. In To appear in Proceedings of ACM SIGCOMM '97, 1997.

.... without sending the static portions, and to allow a compact representation of repetition within a resource (through the use of the LOOP construct) In fact, both of these goals can be achieved through other means: the former via delta encoding [1, 8, 12, 14] and the latter via compression [12, 13]. A fundamental issue with delta encoding is the management of past versions, since a server and Douglis, et al. USITS 97 HPP: HTML Macro Preprocessing to Support Dynamic Document Caching 7 loop FONT SIZE= 1 FACE=ARIAL HELVETICA b var counter ) a href= var queryurl var ....

Henrik Frystyk Nielsen, James Gettys, Anselm BairdSmith, Eric Prud'hommeaux, Hakon Wium Lie, and Chris Lilley. Network performance effects of HTTP/1.1, CSS1, and PNG. In Proceedings of SIGCOMM'97, pages 155--166, Cannes, France, September 1997. ACM.

No context found.

Henrik Frystyk Nielsen, Jim Gettys, Anselm Baird-Smith, Eric Prud'hommeaux, Hokon Wium Lie, and Chris Lilley. Network performance effects of HTTP/1.1, CSS1, and PNG. In ACM SIGCOMM Symposium on Communications Architectures and Protocols, Cannes, France, September 1997.

No context found.

Henrik Frystyk Nielsen, James Gettys, Anselm Baird-Smith, Erick Prud'hommeaux, Hakon Wium Lie, and Chris Lilley. Network Performance Effects of HTTP/1.1, CSS1, and PNG. In In Proceedings of SIGCOMM'97 Cannes, France, pages 155--166, September 1997.

No context found.

Henrik Frystyk Nielsen, Jim Gettys, Anselm Baird-Smith, Eric Prud'hommeaux, Hakon Wium Lie, and Chris Lilley. Network Performance Effects of HTTP/1.1, CSS1, and PNG. In ACM SIGCOMM Conference, 1997.

First 50 documents

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

CiteSeer.IST at NUS - Copyright Penn State and NEC. Hosted by the School of Computing, National University of Singapore.