Relationship between SNRM, error rates, and throughput.

[#1]

Does anyone know what the relationship is between error rates and actual throughput (i.e. is it possible to quantify at what point the losses due to correcting or re-transmitting data outweigh the increase in connection rate due to having a lower SNRM?)
I ask this because the DLM refuses to select anything other than a 3dB margin on my ADSL2+ connection (it even applied a low level of interleaving to presumably "justify" its decision.) At that level it generates a large number of errors, particularly in the evenings when the margin drops to around 1dB or even less.
I now run it "tweaked" to a 5dB target because it would occasionally drop the connection altogether and it is perfectly stable, but I might put up with the occasional drop if I thought the speed gains outweighed the losses.

[#2]

Well suppose one in 10 of your received packets contained an uncorrectable error and required retransmission and one in 10 of the retransmitted packets was also errored.  You would end up with a bit less than 90% of your nominal bandwidth, which would not be too severe.  On the other hand if 5 in 10 of your received packets contained an uncorrectable error and 5 in 10 of the retransmitted packets needed to be re-retransmitted etc. then you would find your effective bandwidth was far below the nominal value.
As a practical test, visit one of the broadband speed tester web sites at different times of day.  You can then see if a 1 dB noise margin has an adverse effect on the results you achieve.  I advise running a series of test rather than relying on a single result.