This is a very complex topic and an extremely broad question. I'll try a rough answer.
Generally, first you classify your traffic into e.g.
- low-latency (forward at the highest priority)
- don't-drop (latency is no big issue but don't drop anything unless required for low-latency)
- don't-care (do your best but drop if necessary)
Once you've done that, you identify your network hardware capabilities. Most device have multiple priority queues, so you configure each queue to take care of one or more of your traffic classes. You can do that be either using source or destination IP addresses, protocols, source/destination port numbers VLANs, or by assigning an already present traffic class from the packet's ToS field or the frame's 802.1Q tag.
In addition you can configure limits for each traffic class, so that (fake) VoIP class traffic doesn't blot out everything else.
Priorization and limitation is also used for traffic shaping which is another large topic.
Also, a given network may not be able to handle your requirements reliably, especially when aggregated links are not wide enough. The QoS requirements should always be reviewed together with the entire network design.
This is the short answer, there are many books written about these topics.
However, since you're referring to trading application requirements - this isn't usually handled by QoS alone but you need to optimize your entire environment for this purpose as well: select low-latency switches, routers and physical layer transports (even cables), allocate ample bandwidth and optimize the topology for short and quick paths.