Today I joined the ‘shipping’ checking group, instigated by Nat Welch. The idea is to work towards actually shipping something, which I really want to get better at while at RC.
I decided on a simple utility to list IP addresses. This came out of my writing
ifconfig works. Mark Dominus said a
utility that just listed IP addresses would be handy, and he’s never come
across one. I did a little poking around the googles and stackoverflows, and
most answers involved
grep, and either
awk. A few
listed the Linux only
ip addr show which still has a lot of cruft in its
output. I also came across another Linux only option,
--all-ip-addresses, which prints them one per line, ignoring loopback and link
The last one is pretty close to what I would like, but it’s missing some handy flags to specify IPv4 or IPv6, which interface, or whether to include link local or loopback addresses. It’s also Linux only, which a standalone utility could work around. I’m aiming to have a fairly polished first version for Linux systems this week, possibly even dabbling in Debian packaging.
Of course, right after deciding that, I got distracted following up on some links Anil Madhavapeddy gave me last Thursday while he was at RC. It turns out the IPC performance is of a lot of interest to virtualization people, and he had this big set of benchmarks. I also read through slides from his FOSDEM 2012 talk entitled The Wild West of UNIX I/O. This goes into a lot of details around non-uniform memory architecture (NUMA) and how it affects latency and throughput between core in a many core machine. It’s incredibly fascinating that the interconnections between the cores can be inferred from measuring IPC performance.
Even more interesting to me was the little bit at the end on Fable I/O, an
‘ongoing attempt at a “new” sockets API for high-performance data’. Sadly, the
‘ongoing’ part of this no longer seems valid: the best I could find was a
libfable repository that hasn’t seen any updates in a couple of
years. This got me a bit excited, and I started reading up a bit more on NUMA
support in Linux. I came across a Linux Weekly News article
from 2012 on NUMA aware scheduling support. So many things to think about:
- the idea of home NUMA nodes for processes
- moving physical pages between nodes while leaving virtual addresses intact
- grouping processes together into NUMA groups that will always share a home node
- the performance gains possible by understanding and exploiting hardware details
On the performance note, a quote from the LWN article stands out:
Without the NUMA balancing patches, over time, the benchmark ended up with just as many remote memory accesses as local accesses - allocated memory was spread across the system. With the NUMA balancer, 86% of the memory accesses were local, leading to a significant speedup.
NUMA: serious business.
After a while I was able to settle down again from all this IPC excitement and
think about how to approach the list-IP-addresses utility. I decided to go with
some README-driven development, and drafted up a README. After a bit,
I changed the aim to writing a manpage. This was driven by a secret desire to
use AsciiDoc (see also: AsciiDoctor), as I’ve been wanting to check it out as a Markdown alternative
for a while. I knew it had good support for manpage generation. In fact, the
git docs are in AsciiDoc.
So, current status:
- I have some flags being parsed
- can list IPv4 addresses
- manpage needs some rewording and reworking
- the tool needs a name. I had
lsof, but it looks wrong. I’m now thinking
lsaddr. I am bad at names.