Path: chuka.playstation.co.uk!scea!greg_labrec@interactive.sony.com From: "Wayne K. Werner" Newsgroups: scea.yaroze.freetalk Subject: Re: Audio breakup in the auditorium Date: 31 Aug 1997 07:34:44 GMT Organization: WermX Software Lines: 56 Message-ID: <01bcb5e0$2da633e0$9bbf43ce@wkwerner> References: <01bcb55b$48b98360$6cbf43ce@wkwerner> <340850DD.92B0A366@msu.edu> NNTP-Posting-Host: port55.con2.com X-Newsreader: Microsoft Internet News 4.70.1161 Charles Henrich wrote: > There is no such thing as a "natural" internet packet, in fact you will > find the vast majority of any bulk data transfers are always 1500, to > limit the header overhead associated with each packet. You are incorrect there. The packet size used in the internet, and by internet routers is indeed 576 bytes. When you use larger packets, they must be disassembled into smaller packets on send and reassembled on receive. The size of the receive window, also to prevent fragmentation, must be a multiple of the data size, which is 576 - 40. Yes, you are adding "extra" bytes up front. But the extra bytes are added anyhow when a 1500 byte packet is fragmented, with the further inconvience of not dividing up neatly, causing the system to wait for a second 1500 byte packet to fill up the space after any leftover data from the first packet. > Im curious as to how you measure this, downloads should be *slower* with Don't take my word for it, try it! FTP down a file from a nearby site, time it, make the fix, and download the same file. Everyone I know who has made the change has experienced a great increase in speed, all claiming at least 2 times, some claiming as much as 3 or 4. I myself have experienced an increase of at least 2 times. > What shrinking the MTU size WILL do for you is make interactive response > during high bandwidth usage feel better. Your keypresses get more > chances at being sent off when your system is generating a new packet more > frequently than at 1500. Keypresses and such are processed locally. The size of a packet has no influence over when packet sending is initiated. For example, I compose a message locally for the newsgroup. When I press the post to news group button, my message is assembled into packets and sent, whether it's 20 packets full of data or one packet with a single byte of data in it. What does matter is whether or not I am exploiting the "natural" size of internet packets or not. This is not a perceived increase because of increased interactivity, but a real improvement of transfer speed. Again, don't take my word for it. Experiment for yourself (and let me know the results, if you don't mind). Also, don't forget... just setting the MTU is insufficeint to realize the improvements. The receive window must also be set at n * (576 - 40). experiments with different values for n, but 4 appears to be optimal for most people. (40 is the number of bytes in the packets header). The problem with using larger values with rwin is that the entire buffer is requested for resending if any one packet comprising it is damaged in transmission. If you have an extremely clean connection consistently, you may find better results with values of 6 or 8. Again, experiment. -- Wayne K. Werner wkwerner@con2.com