Although this is described in stupendous details elsewhere and in our manpages, this question gets asked a lot and happily there is a simple answer that does not need full comprehension of traffic control.
This three line script does the trick:
tc qdisc add dev $DEV root handle 1: cbq avpkt 1000 bandwidth 10mbit tc class add dev $DEV parent 1: classid 1:1 cbq rate 512kbit \ allot 1500 prio 5 bounded isolated tc filter add dev $DEV parent 1: protocol ip prio 16 u32 \ match ip dst 195.96.96.97 flowid 1:1
The first line installs a class based queue on your interface, and tells the kernel that for calculations, it can be assumed to be a 10mbit interface. If you get this wrong, no real harm is done. But getting it right will make everything more precise.
The second line creates a 512kbit class with some reasonable defaults. For details, see the cbq manpages and Chapter 9.
The last line tells which traffic should go to the shaped class. Traffic not matched by this rule is NOT shaped. To make more complicated matches (subnets, source ports, destination ports), see Section 9.6.2.
If you changed anything and want to reload the script, execute 'tc qdisc del dev $DEV root' to clean up your existing configuration.
The script can further be improved by adding a last optional line 'tc qdisc add dev $DEV parent 1:1 sfq perturb 10'. See Section 9.2.3 for details on what this does.