Thanks a lot for the clarification. I did read the docs but the part with the sendfile and no successive writes was what i did not connect. I thought that because of "not for the
transmission of the whole response" it didn't apply in that case.

I tried to fix that by adding 'sendfile_max_chunk 32k' but this didn't help. To me when doing that it seems to be weird that many timers are added and deleted when starting the transfer

2016/01/30 12:54:11 [debug] 9259#0: *1 event timer del: 12: 1454158511891
2016/01/30 12:54:11 [debug] 9259#0: *1 event timer add: 12: 1:1454158451892
2016/01/30 12:54:11 [debug] 9259#0: *1 event timer del: 12: 1454158451892
2016/01/30 12:54:11 [debug] 9259#0: *1 event timer add: 12: 1:1454158451894
2016/01/30 12:54:11 [debug] 9259#0: *1 event timer del: 12: 1454158451894
2016/01/30 12:54:11 [debug] 9259#0: *1 event timer add: 12: 120000:1454158571894
2016/01/30 12:54:11 [debug] 9259#0: *1 event timer del: 12: 1454158571894
2016/01/30 12:54:11 [debug] 9259#0: *1 event timer add: 12: 1:1454158451922
2016/01/30 12:54:11 [debug] 9259#0: *1 event timer del: 12: 1454158451922
2016/01/30 12:54:11 [debug] 9259#0: *1 event timer add: 12: 1:1454158451923
2016/01/30 12:54:11 [debug] 9259#0: *1 event timer del: 12: 1454158451923
2016/01/30 12:54:11 [debug] 9259#0: *1 event timer add: 12: 1:1454158451924
2016/01/30 12:54:11 [debug] 9259#0: *1 event timer del: 12: 1454158451924
... (many more follow)

So the only way to fix that for the moment to me is to set 'sendfile off' and also increase the 'send_timeout' to a more reasonable value in that case. Again thanks a lot for the clarification!

Cheers,
Markus