No. At least, not in the case of graph computation. This was the conventional wisdom with some people, but circa a year or so ago when we did some measurements, no.

https://github.com/frankmcsherry/blog/blob/master/posts/2015...

https://github.com/frankmcsherry/blog/blob/master/posts/2015...

If you don't want to read everything (but you should, before trusting the conclusions), the relevant measurements are here:

https://github.com/frankmcsherry/blog/blob/master/posts/2015...

Edit: Reading the paper, they cite these measurements and present their work as "only [..] 8.8× slower than a distributed in-memory" system (GraM, a different system than timely dataflow). The quote is on the first page.