If the desired load on the system increases, the throughput ideally increases as well. The theory of constraints is based on the assumption that there will be at least one constraint that will throttle the system's throughput. The constraints or bottlenecks therefore cause a performance regression.
Like a chain that is only as strong as its weakest link, the constraining resource limits the overall performance of the system or certain functionality thereof. It prevents the application from handling more load while other resources are not fully utilized. Only by increasing the flow through the constraining resource, that is, removing the bottleneck, will the throughput be increased. If the system is optimized around the bottleneck rather than removing it, the responsiveness of the overall system won't improve and, ultimately, it may even decrease.
It's therefore crucial to identify what the bottleneck is. The overall performance won't improve, before the limiting bottleneck gets targeted.
For example, throwing more CPU power at an application with high CPU utilization probably won't help to achieve better performance. Maybe the application isn't performing well because of other root causes than insufficient CPU.
It's important to mention here that the limiting constraint likely is external to the application. In a single, monolithic application, this includes the hardware and the operating system, with all running processes. If other software running on the same hardware heavily utilizes the network adapter, the application's network I/O and overall performance will be affected as well, even if the root cause, the limiting constraint, isn't the in responsibility of the application.
Inspecting performance issues therefore needs to take more into account than just the application itself. The whole set of processes running on the same hardware can have an impact on the application's performance, depending on how the other processes utilize the system's resources.
In a distributed environment, performance analytics also involves all interdependent applications, that interact with the application, and the communication in between. In order to identify the constraining resource, the overall situation of the system has to be taken into account.
Since the applications are interconnected, improving the responsiveness of a single system will affect others and can potentially even decrease the overall responsiveness. Again, trying to improve the wrong aspect, such as optimizing around the bottleneck, will not improve rather than most likely even downgrade the overall performance. An application that connects to an external system that represents the bottleneck , puts certain pressure on the external system. If the application's performance is tuned, rather than the external application, the load and pressure on the latter is increased which ultimately leads to overall worse responsiveness.
In distributed systems, the situation with all interdependent applications involved vastly complicates solving performance issues.