My biggest concern -- the same concern I raised a few months ago -- is the message of the paper. Part of the paper is "here is a broad survey of approaches to graph computation", which, while interesting in general, is not going to be of particular interest in an academic conference that's not focused on surveys. The other part of the paper is presumably focusing on results. My concern here is that the current approach appears to be "A is faster than B" but not "A is faster than B and this is why and here's the lesson we learn from that". There is little analysis in the current draft, and it's already over 9 pages with sections missing, and without significant analysis and some real take-home messages from the paper I don't think that a conference will be particularly interested. Doing all this benchmarking work has real potential, but it has to be coupled with analysis of the results.
My concern hasn't changed. I do like the very last paragraph in section VI. That has some insight (that performance differences are in part due to kernel count). It is a qualitative rather than a quantitative analysis; having numbers would really help, and some microbenchmarks that show kernel-launch cost on both card types. But this does seem contain some real insight that is unfortunately lacking from the analysis in many of the other parts of the paper.
Same thing I said in May still applies:
My concern hasn't changed. I do like the very last paragraph in section VI. That has some insight (that performance differences are in part due to kernel count). It is a qualitative rather than a quantitative analysis; having numbers would really help, and some microbenchmarks that show kernel-launch cost on both card types. But this does seem contain some real insight that is unfortunately lacking from the analysis in many of the other parts of the paper.