(Note: We discuss some network-related considerations on the "Ask the Expert" call, the replay link is elsewhere on this page).
Certainly, there are differences for cloud-specific workloads vs a dedicated server (see 'auditing' link below). But I'd suggest that it might be more important to consider where a load test originates vs where a target system/application is running. Both can be on dedicated servers, either could be cloud-based, or both could be cloud-based. For internal services (whether on private cloud or dedicated servers) internal load generation will have the lowest latency --- and the latency will be more consistent for the duration of a test. There is, as you imply, some impact from the presence of cloud - this is discussed further at the auditing link, below (though it targets 'virtualization' as a topic, clouds share the same factors).
For public cloud-based applications - whether on OpenStack, Softlayer bare-metal servers, or PaaS clouds like Bluemix, there will be larger latency with less consistency, as the internet will be used to carry traffic. The goals of public cloud-based load generation often include the ability to generate traffic from globally-distributed locations, which implies that latency is not as important in this situation.
Luckily, Rational Performance Tester includes special capabilities that help to minimize the impact of choosing to test over the internet. For more information, please see "Auditing measurement accuracy" at http://www.ibm.com/support/knowledgecenter/SSMMM5_9.0.0/com.ibm.rational.test.lt.doc/topics/c_audit_meas_accuracy.html
Bill Hudacek
Service/API Testing and Service Virtualization, Performance Testing - "All Things Testing"
IBM
These opinions are my own, not those of my employer.