Andrzej,
I think I really need to provide more details here, just as a sample: Ted Rogers is ISP, and he has 8000Mbps synchronous connection to UUNet. His hardware allows to remember (due to RAM and CPU limitations) up to 1,000,000 of IP addresses, and 20,000 TCP ports for each "handshake". And his hardware randomize bandwidth evenly between 1,000,000 x 20,000 = 20,000,000,000 TCP connections. Why evenly? Because it is cheapest solution, no CPU required, no network latency. So, for instance, you use more bandwisth if you use more TCP connections (because connection to UUNet backbone is shared between many users). Now, consider big building, and Router on the roof, which allows only 1024 TCP sessions... If you are using 512 TCP threads, you are guaranteed to use at least 50% of the total bandwidth of shared channel (if your "last mile" allows it). This is the case when 100Mbps "last mile" is dedicated, and 1000Gbps "before last mile" is shared between 256 users with 65536 sessions limitation for all of them. ISP's employees usually don't know such details. They are "help-desk", and they usually ask "Could you please use less than 50Gb download per month? What is your IE version?!" So, suggestion to Crawlers: 1. Consider education of ISP's employees 2. Decrease number of concurrent "alive" threads (aka concurrently open TCP sockets) 3. Increase bandwidth -----Original Message----- From: Andrzej Bialecki Fuad Efendi wrote: > For ISPs around-the-world, thew most important thing is the Number of Active > TCP Sessions. > This is completely false. Having worked for an ISP I can assure you that the most important metric is the amount of traffic, and its behavior over time. TCP sessions? We don't need no stinking TCP, we route good ol' IP ;) Please check your facts before claiming something about all ISPs around the world. -- Best regards, Andrzej Bialecki <>< ___. ___ ___ ___ _ _ __________________________________ [__ || __|__/|__||\/| Information Retrieval, Semantic Web ___|||__|| \| || | Embedded Unix, System Integration http://www.sigram.com Contact: info at sigram dot com
