150 likes | 375 Views
Success disaster for both the Internet and the Web. Conservatively, we need to scale the Internet/Web 104 higher, even ignoring continuous media ...
E N D
Slide 1:The Web versus The Internet Presentation for Joint APPS/transport BOF Los Angeles IETF3/4/96, revised 3/11/96
Jim Gettys jg@w3.org World Wide Web Consortium Digital Equipment Corporation
Slide 2:Mutual Coexistence
The Web is causing the Internet to succeed We dont want the Internet to die of the Webs success
Slide 3:Outline of this Talk
Problems Solution(s) Call to Action
Slide 4:Todays Situation
Congestion on many links (particularly trans-oceanic) Success disaster for both the Internet and the Web Conservatively, we need to scale the Internet/Web 104 higher, even ignoring continuous media I dont believe there is a single magic bullet to solve the problem
Slide 5:How Did We Get Here?
HTTP protocol Feature: simple Bug: simple Fetches single URL per TCP connection Mean size is only a few thousand bytes Bimodal size of URLs, usually short
Slide 6:Consequences of HTTP simplicity
Servers have thousands of connections in time_wait state e.g.. AltaVista server is at > 4 million connections/day, or 50/second averaged over 24 hours Cost is primarily memory, on systems running reasonable TCP implementations Losing congestion information on each URL Poor user perceived performance Netscape multiple connection hack Vanity servers with HTTP result in big servers using 100s of I.P. addresses, due to oversight in protoocol Connection opens may be congesting low bandwidth links, due to lack of flow control Size of routing tables/routing caches in routers?
Slide 7:Available Options
Fix HTTP Replace HTTP 1.X with HTTP-NG Deploy caching Deploy T/TCP Replace TCP with other transport protocol(s) Fix routers to police flows???
Slide 8:HTTP Fixes Required
Persistent connections Caching somewhat buggy, due to problems in 1.0 protocol Multiplex the connection to help time to render Host name must accompany request to solve vanity address problem
Slide 9:Netscape Hack
Problem: Rendering a page requires meta-data of objects embedded in that page Solution (according to Marc A.): Open N connections simultaneously (by default 4) Result: Faster time to render Self congestion of nearly simultaneous opens Unfair use of bandwidth (problem of the commons) Has made a bad problem much worse Even with persistent connections, we will have fairness problems for clients that use multiple connections, versus well behaved clients using one. RED algorithm in routers may help.
Slide 10:HTTP-NG
Preliminary results of (essentially) HTTP 1.X over SCP (mux layer) are very encouraging Simon Spero and Andy Normans NG proxy showing good results on transatlantic links (HTTP 1.X encapsulated in NG over SCP) Finishing design will take quite a while Not a short term solution, in my opinion
Slide 11:Persistent connections
By itself, I dont think they will succeed against Netscape hack, as it hurts Time to Render Need multiplexing layer to solve time to render problem (have preliminary design in hand) Prototyping mux layer for HTTP 1.x right now, with provisions for multiple protocols over the same connection, to allow graceful NG transition Transition to a MUX layer has all the same deployment problems of NG, but requires much less work than NG for clients and servers to implement Mux layer can be done quickly
Slide 12:Caching
Needs both HTTP fixes and deployment Deployment by ISPs Benefits: Helps perceived response Gives appearance of more robust system Problems: Dont expect too much: 60-70% hit rate, but is affected by demographics problems How do you find your cache? Need to be able to autoconfigure caching, due to naďve users Administration hassles?
Slide 13:T/TCP
Would help by: Reducing packet exchanges on opens (which are not flow controlled) Reduce latency to end user (bored users == hitting stop buttons on browsers) Is a better match for current HTTP than once persistent connections are deployed A net gain in any case Concerns: uses some state between connections Must understand consequences on big Web server systems Must fully understand congestion consequences, and potential performance cliffs Unlikely to help in short term due to deployment time
Slide 14:No Magic BulletShort Term Fixes
Fix HTTP-1.0 Vanity addresses Caching bugs Persistent connections Should be minimum set of deliverables for HTTP 1.1 Add MUX layer as soon as possible otherwise I expect persistent connections will not succeed in the market Encourage deployment of caching servers at ISPs Fix economic bugs in Internet traffic Deploy and enable RED algorithm in routers, to help fairness problem
Slide 15:No Magic BulletLonger term
Finish NG work (will help bandwidth and latency over high latency connections) Deploy and use T/TCP for the web Multicast transport with caching might solve flash crowd problems Need better name services D&B has 40 million businesses How to find nearest cache!?!? Flow policing in routers???? Get Internet transport, Web, and distributed systems people working on longer term solutions