Rephrase the text on perf gains of high end boxes, thanks @steveloughran

input/1_intro.md

@@ -15,11 +15,11 @@ However, few people have infinite resources. Hence, they have to find the right
 
 It is a current reality that the best value is in mid-range, commodity hardware - as long as the maintenance costs can be kept down through fault-tolerant software.
 
-Computations primarily benefit from high-end hardware to the extent to which they can replace slow network accesses with internal memory accesses. In tasks that require only small amounts of communication between nodes, the performance advantage of high-end hardware is limited.
+Computations primarily benefit from high-end hardware to the extent to which they can replace slow network accesses with internal memory accesses. The performance advantage of high-end hardware is limited in tasks that require large amounts of communication between nodes.
 
 ![cost-efficiency](images/barroso_holzle.png)
 
-Even in tasks that require high amounts of communication, the performance gap between high-end and commodity hardware decreases with cluster size assuming a uniform memory access pattern across all nodes (e.g. the figure above from [Barroso, Clidaras & Hölzle](http://www.morganclaypool.com/doi/abs/10.2200/S00516ED2V01Y201306CAC024)).
+As the figure above from [Barroso, Clidaras & Hölzle](http://www.morganclaypool.com/doi/abs/10.2200/S00516ED2V01Y201306CAC024) shows, the performance gap between high-end and commodity hardware decreases with cluster size assuming a uniform memory access pattern across all nodes.
 
 Ideally, adding a new machine would increase the performance and capacity of the system linearly. But of course this is not possible, because there is some overhead that arises due to having separate computers. Data needs to be copied around, computation tasks have to be coordinated and so on. This is why it's worthwhile to study distributed algorithms - they provide efficient solutions to specific problems, as well as guidance about what is possible, what the minimum cost of a correct implementation is, and what is impossible.
 

input/6_appendix.md

@@ -4,7 +4,7 @@ If you've made it this far, thank you.
 
 If you liked the book, follow me on [Github](https://github.com/mixu/) (or [Twitter](http://twitter.com/mikitotakada)). I love seeing that I've had some kind of positive impact. "Create more value than you capture" and all that.
 
-Many many thanks to: logpath, alexras, globalcitizen, graue, frankshearar, roryokane, jpfuentes2, eeror, cmeiklejohn, stevenproctor and eos2102 for their help! Of course, any mistakes and omissions that remain are my fault!
+Many many thanks to: logpath, alexras, globalcitizen, graue, frankshearar, roryokane, jpfuentes2, eeror, cmeiklejohn, stevenproctor eos2102 and steveloughran for their help! Of course, any mistakes and omissions that remain are my fault!
 
 It's worth noting that my chapter on eventual consistency is fairly Berkeley-centric; I'd like to change that. I've also skipped one prominent use case for time: consistent snapshots. There are also a couple of topics which I should expand on: namely, an explicit discussion of safety and liveness properties and a more detailed discussion of consistent hashing. However, I'm off to [Strange Loop 2013](https://thestrangeloop.com/), so whatever.
 

readme.md

@@ -31,7 +31,7 @@ You need to install Calibre first for the HTML to epub/mobi conversion.
 
 # Thanks
 
-Many many thanks to: logpath, alexras, globalcitizen, graue, frankshearar, roryokane, jpfuentes2, cmeiklejohn, stevenproctor and eos2102 for their help!
+Many many thanks to: logpath, alexras, globalcitizen, graue, frankshearar, roryokane, jpfuentes2, cmeiklejohn, stevenproctor, eos2102 and steveloughran for their help!
 
 # Licence