EFOSC2 correlated readnoise problems

PESSSTO Night 20131125  (observers : D. Bersier, M. Dennefeld ; data reduction : S. Smartt) 

We have seen this problem intermittently during PESSTO but this night is particularly seriously affected, as it appears to affect at least 4hrs of exposures. This is becoming a significant loss of time on EFOSC2. At least half a nights data has been compromised, some of it is usable, some of it is totally unusable. It is also difficult for observers and TOs to spot the problem, and see the severity. As this has now happened repeatedly (although fewer frames have been lost in the past) PESSTO would like to ask ESO to investigate fully and have a solution during the night if it appears again. A summary for the instrument/technical team (and observer related information) is given below. 

Bias frames at start of night look OK :  EFOSC.2013-11-25T20:05:14.298.fits - EFOSC.2013-11-25T20:10:22.614.fits   

Also, all frames with low background (below about 20 ADU above background) up to frame EFOSC.2013-11-26T00:39:23.441.fits look OK. 

Problem starts to be obvious in 

EFOSC.2013-11-26T01:06:29.294.fits SN2009ip   1800s Gr#11  

and continues to 

EFOSC.2013-11-26T06:04:08.985.fits  (i.e. around 5.5hrs observing, affecting around 25 frames. The higher background imaging frames probably suffer less). 

It does not appear to be stable enough to remove in that division or subtraction of two subsequent frames does not remove the pattern. 

Note for observers and reducers - be careful with checking for this problem, can be missed easily at first sight 

LSQ13cuw has 2x1800s exposures. Both look normal at first glance. But when added together, a correlated read noise pattern appears. White and red are the reduced spectra for each exposure and green is the mean of the two (yscale is normalised and offset). 

And if the two individual spectra are divided the result is as expected if the pattern is correlated noise. The "high" noise regions in the green (added) spectra line show correlated structure which divide out. So they appear to be correlated in the two adjacent frames.