On 3/01/2014 11:45 a.m., Eliezer Croitoru wrote: > Hey Amos, > > There is one exception case which is not exactly like the description. > In a case you have a LAN connection with for example 1Gbps line. > The internet uplink is 20Mbps. > In this particular case when the client is using a Distributed FS or > Distributed requests such as in: > http:/server1.example.com/file1.tar.gz > http:/server150.example.com/file1.tar.gz > > And in the above case the client is requesting from all the servers a > 206 partial content request there is a possibility of accessing > server1.example.com = 192.168.0.100 > server150.example.com = 192.168.0.101 > Domain sharding *is* one of he cases degrading traffic flow via TCP congestion control. It trades resource use in part of the same path for resource use on other paths, AND changes the identifiable destination so smarter congestion control cannot identify and treat the two as related connections. > While it is assumed that the client is accessing some origin server with > some "tricks" it is actually using local cache peers. > It can be assumed that the cache service provider will be using some CDN > network to allow the clients the download of the content in a very fast > speed. These tricks are all added complexity and difficulty for the administrator to manage. They will work or break regardless of the use of Ranges (and probably more often when Ranges are used). > > It is however overloading the TCP level of the connection.. > It can be assumed that it will not cause the upstream service providers > a load in the TCP level. > > To design such a network there is a need in a very large amount of > resources but work of tens of years can achieve such a thing. > > Eliezer > > On 01/01/14 07:27, Amos Jeffries wrote: >> NOTE: download managers which open parallel connections are*degrading* >> the TCP congestion controls and reducing available network resources >> across the Internet. Reducing their parallel requests to a single fetch >> is actually a good thing. > Amos
