I may have finally written up the papers I read some time ago, but
I still have a number of papers from that pile, plus a couple more
recently-added ones. Time to discuss them:
- Mosh: An Interactive Remote Shell for Mobile Clients This
paper reminds me how I don't tend to take an innovative approach. I
thought of ssh as good enough, and am generally suspicious of
reinventing the TCP wheel. Then, this paper shows how you can make
remoting over flakey wireless connections much better if you try.
D'oh. ("mosh" itself is rather nice, but I'd really rather like good
scrollback.)
- Practical Byzantine Fault Tolerance One of the classic
papers on distributed systems. The bit on committing operations is
very plausible, but I look at the section on changing views and... its
correctness is not intuitive to me. Tricky.
- Data Compression Using Long Common Strings Data compression
is often rather tricky, but as you might expect from Bentley and
McIlroy, this is both simple and effective. Fun.
- Large-scale cluster management at Google with Borg And this
one suddenly looks very relevant to my work! The first section or two
seemed a bit vague and naff, but the details fill in over the later
sections to give a reasonable impression of what they had and now
have. The future work section (especially Kubernetes) looks very
interesting, and the references should give me plenty to chew on.
- Breakthrough silicon scanning discovers backdoor in miliary
chip A somewhat melodramatic title sits at the top of a
fundamentally unremarkable paper. Unremarkable, perhaps, but still
fun. The "backdoor" is in an FPGA used in, among other things,
military technology, allowing the bitstream to be read out. Looks like
an undocumented debug feature. The interesting things are a) how the
advertising material lies so strongly about the security b) the
technology used to identify the undocumented feature.
- Trafficking Fraudulent Accounts: The Role of the Underground
Market in Twitter Spam and Abuse Yet another fun little paper
investigating the online criminal economy. (If you like this, go read
all the Silk Road reporting - not directly related, but also
interesting.)
- TCP ex Machina: Computer-Generated Congestion Control
Another paper by the guy who did Mosh, although I came across this one
first. In short, by applying a simulate-and-optimise approach to TCP
congestion control, something can be produced which appears to work
rather better than standard (human-designed) congestion control
algorithms. There's lots of interesting stuff here. Is the evaluation
biased by the fact that it's using models not entirely dissimilar to
the ones used to do the optimisation, leading to over-fitting? Can we
know it's not going to have horrible corner cases because it's not
human-designed? When setting up an optimisation problem, how the
problem is framed will feed into the result, and that's rather
interesting. For example, packet loss is not used. Instead, RTT
variation is used to detect queuing, which is rather nice given the
problem of buffer bloat. All rather cool.
Posted 2015-05-31.