My hypothesis is that you have a single saturated load generator which is slowing to almost zero as you go from 40 users to 100 under truclient. The weight of this virtual user type is why I shy away from its general use and relegate it to a handful or less where there is some sort of "rendering" demand from people who do not understand the focus of performance testing.
Look to your test bed. Minimally you should have at least three load generators in your test bed, hardware matched. Two for primary load and one for a control set. This is independent of the controller. Run the majority of your load on the two "primary load" generators and one virtual user of each business function type on the control generator. Watch your transaction response times during the test
If you see the transaction response times from your control group begin to vary from the global group you have a problem. Global getting longer on average while the control group continues at the same pace or gets faster is an absolute sign of load generator distress, independent of tool and independent of protocol used for testing.
Also consider moving to a model where the majority of your load is http and only a few are Truclient, where they are being run to satisfy a requirement which cannot be satisfied any othere way. There may be technical environments which demand a higher layer development method, but these are few and far between, especially when you remove the calls to third party services that you would not include in your performance test anyway.