Sanjeev Gupta via devel <devel@ntpsec.org>: > This is a follow on to Eric's email a few hours ago, I am keeping that > thread clean. > > (The last 3GL I programmed in was Fortran, and not the 77 version. I can > read bash scripts and C pseudo-code) > > The literature I can find speaks of Go GC being improved in 1.5, such that > the STW phase (the "sweep") is now less than 1ms. This is impressive, but > for NTP, this places a lower bound on our jitter. > > What am I missing?
Well, first, the historical target for accuracy of WAN time service is more than an order of magnitude higher than 1ms. The worst-case jitter that could add would be barely above the measurement-noise floor at worst, and more probably below it. Second, Go 1.5 was a long time ago. STW pauses are much shorter now. The graph at https://github.com/lni/dragonboat indicates that even under heavy load STW in GO 1.12 never went above 600 microseconds and is usually somewhat below 400. We can expect this figure to decrease rathrr than increase in the future - reducing latency spikes is high on the Go development team's objectives. Third, most of the code isn't stall-sensitive at all. There are a couple of critical regions that need to be guarded, which I think we can accomplish by eirther (a) turning GC off before we enter them and turning it on again after, or (b) some Lampson-like tricks for tetecting when the interval was interrupted by GC and discarding any resulting sample. I don't think we'll ever need to go to that third level, but we can deal with it if we need to. -- <a href="http://www.catb.org/~esr/";>Eric S. Raymond</a> _______________________________________________ devel mailing list devel@ntpsec.org http://lists.ntpsec.org/mailman/listinfo/devel
