This is an excellent paper reporting on the experience of implementing XIOS, an IO server into the ECMWF intetgrated forecast system.  The paper has a good introduction and motivation with well presented results and discussion. I can't remember the last time I read something on this sort of topic which was as easy to consume. I find myself with little to contribute in terms of suggested modfications!

I recommend publishing as is, but recommend addressing a few minor points during the final submission process.  That said the authors and editors might consider whether section 6 and the associated figures are really necessary? It's not really about choices around XIOS, but about the difference between GRIB and NetCDF ... it's a different topic ...

My list of minor points:

- obsolete not obsolescent (in the abstract, it is correct in the text)

- I would recommmend adding the word "existing" before  "IFS I/O server" at the end of page 1.

- line 5 page 2. I would discuss increasing spatial resolution, not spatio-temporal resolution (even if the timestep has changed).

- For the caption of figure 2, I think it would be helpful to expand a little bit more about the meaning of the axes in this trace. While it is easy to infer, it takes a few minutes for those who are not familiar with the output. I realise this is discussed further in B1, but frankly most readers might not get to the appendix!

- the first sentence of the last para of page 9 is hard to construe given the previous para talks about "the three tests" and then this para talks about "the three MIP configurations).  After two or three reads and a look at the table I understood it, but it could be cleaner.

Overall an excellent paper.   On the issue of scientific reproducibility, that the IFS parallel IO scheme is not publically available makes this aspect of the paper difficult or impossible to reproduce. 

I only have a couple of suggested corrections:

Page 13, line 14: "where" should be "were"

Page 17, line 28: "of" should be "for"

Page 17, line 29: " will be blocking the model time stepping in doing so" should be "will block progress of the model time step in doing so"

I would also like to point out newer reference documents for the parallelIO library: https://github.com/NCAR/ParallelIO#references