ECOS fails but CVXOPT doesn't in SOCPs

[Heikki Arponen]

This happens for quite large dimensions, about 1000 or so and in many situations, mostly in quadratic programs... I usually get 

" No further progress possible, recovering best iterate (40) and stopping.
NUMERICAL PROBLEMS (reached feastol=7.8e-04, reltol=1.8e-01, abstol=1.5e-01)."

It seems to converge at first, but then the gap starts increasing again... if I lower the dimension, ECOS eventually converges. Maybe I should scale my variables differently?

This happens in several QP programs, where I use e.g. quad_form so that CVXPY should convert that to SOCP automatically.

Anyone know what would usually cause this?

[Steven Diamond]

Could you post this and your code as an issue on the ecos github? Alex Domahidi (the author of ecos) is good about analyzing numerical issues like this.