Talk:TCP tuning

This Page
This page started from a workshop held by Internet2 on "Bridging the Wizard Gap" in Ann Arbor, MI.
 * Very interesting topic, I'd love to see it grow 84.47.47.221 00:06, 14 January 2006 (UTC)

Proposed Changes
I think it'd be best to separate the definition of bandwidth-delay product (BDP) topic of TCP Tuning. I'd recommend cleaning up the paragraph on BDP and then making a new entry for it. This should be linked from the BDP disambig. page and a fuller description of TCP tuning.
 * Note Bandwidth-delay product exists. BTW, please sign your comments. --QEDquid 08:01, 23 January 2006 (UTC)
 * Please refer to Talk:Internet_protocol_suite... I added some information to that talk page because I couldn't find this page or the Rwin page very quickly. Wisepiglet 05:46, 7 December 2006 (UTC)

packet loss throughput limit
I've played a bit with the quoted formula and this simply cannot work. I don't want to suggest Mathis et al are wrong, but surely this has been quoted out of context. My assumption: Zac67 (talk) 15:29, 29 July 2009 (UTC)
 * gigabit ethernet
 * one switch (store & forward), two hosts
 * MTU is 1500 bytes = 12000 bits (ACK packet is also 1500 bytes for simplicity)
 * Packet loss is 10^-6
 * Minimum RTT is 0.048 ms (12000/10^9 * 4 - switch requires zero time for MAC lookup and CRC verification)
 * Throughput limit should be 175 mbps ~= 21.9 MB/s - this is complete nonsense as can easily be observed. If you think that 10^-6 is a very pessimistic packet loss: while sending 83000 packets/s it means you lose one packet every 12 seconds. Why should this limit your effective throughput?

I agree - the formula is obviously misquoted. On a perfect link with loss probability almost equal to zero, the throughput diverges against infinity! I'm removing the formula; if somebody knows more about the matter, feel free to include a more plausible one. Zxb (talk) 11:16, 4 February 2012 (UTC)

The formula of course is for an additional limitation for the bitrate. It also applies only in the congestion avoidance mode, so it requires a reasonable amount of packet loss. Adding the formula back with more about its application. Alinja (talk) 16:47, 10 May 2012 (UTC)