High-frequency trading is the primary application, where 5ns can represent millions in profit as firms compete to execute trades first, but you'll also see benefits in distributed database synchronization, real-time financial risk calculations, and some specialized scientific computing workloads.
High Frequency Trading is one.
Anything else? Because that's the only one I can think of.
I'd expect HPC would be another, since a lot of algorithms that run on those clusters are bottlenecked by latency or throughput in communication.
For the parent: and not only bottlenecked at single hops but also hampered by the propagation of latency as the hops increase, depending on the complexity of the distributed system design.
> […] by the propagation of latency as the hops increase […]
Which is why you get network topologies other than 'just' fat tree in HPC networks:
any high-utilization workload with a chatty protocol dominated by small IOs such as:
* distributed filesystems such as MooseFS, Ceph, Gluster used for hyperconverged infrastructure.
* SANs hosting VMs with busy OLTP databases
* OLTP replication
* CXL memory expansion where remote memory needs to be as close to inter-NUMA node latency as possible