The [TCP/IP] stacks in [Linux] and most other recent OperatingSystem~s use a clever system called __path [MTU] discovery__ to address this problem. The [MTU] for a new connection is set low, but gradually increases, while the packets are sent with the __don't fragment__ flag set. When the [MTU] grows beyond the smallest [MTU] of any link anywhere in the route, the responsible gateway will refuse to route the packets because they're too large to send them without fragmenting them but it's not allowed to do that. Instead, it will notify the source host with a __Need to fragment, but don't fragment is set__ [ICMP] message. At this point the sending host knows the maximum fragmentation-free [MTU] size it can use on this connection, which is the most bandwidth-efficient size for packets. 

  

 Unfortunately, some particularly clueless SysAdmin~s ignorantly block any and all [ICMP] packets from passing through their gateways without fully understanding why [ICMP] is an important part of [IP], and/or without thinking of the consequences. Banks seem to be a major transgressor here. 

  

 See also: 

 * Some common (and not so common) [MTU sizes | http://www-12.lotus.com/ldd/doc/domino_notes/5.0/readme.nsf/0/f397306f052d9ea3852567740049a10a?OpenDocument]