Today we shipped NTPsec 1.0.
Cleaning up the Classic codebase was a lot of work. It’s taken us two and a half years of effort to get it as hardened, cleaned, and stripped down as we think it needs to be for secure operation.
In the process, we’ve improved its timekeeping accuracy. We’ve added major new features like the ability to run autonomously, using only local clock sources with no network peers. We’ve greatly improved the monitoring and statistics tools. We’ve even cleaned up the notoriously rebarbative configuration syntax.
Perhaps most importantly, we’ve removed 76% of its original bulk, over 178 KLOC down to 55KLOC. That huge reduction in attack surface has dramatically reduced our vulnerability to exploits. Since the fork date we have closed off potential exploits so effectively that we avoided more than 75% of the CVEs issued against Classic, and there have been zero NTPsec-unique CVEs.
In keeping with our focus on security, we are actively working with the IETF on the forthcoming NTS (Network Time Security) standard. We expect to be first or second interop on that.
We believe we can make significant improvements in startup performance, decreasing time to clock synchronization and reducing the odds that ntpd might transiently sync to a bad source before being slowly corrected by other peers. This will have a practical impact on how soon NTP-using systems are fully ready for production after boot.
The full-autonomy feature introduced in our 1.0 is important for sites with very stringent security requirements that want to avoid punching a hole in their firewall in order to reach network check peers. As now shipped it only works with a subset of our reference-clock drivers. That can be fixed.
There may be things to be done that can reduce ntpd’s vulnerability to GPS clock rollovers.
We think there may be problems with cross-era compatibility, and while the 2036 era 1 turnover is still 19 years off it’s not actually too soon for some really tough verification and testing.
There’s been so much to do cleaning up and hardening the plumbing around the core time-sync algorithms that we have not tried to improve those yet. There are possibilities…
In the longer term, the Network Time Protocol is showing its age. It was designed when network bandwidth was far more expensive than today, and trades away some desireable gains (like not being susceptible to era wraparound) in order to make its packets small. It also needs redesign to be fully IPV6-ready. NTPv5 is a big job, but somebody needs to do it and that will probably be us.