The Next Level 2

One idea for improving time stability was to remove one suspected source of the problem. The predominant theory for the cause of most of the instability in conventional ntp time servers is jitter and drift in their quartz timebases.

Even with a GPS Clock for synchronization, it is still a low-precision quartz oscillator that you are trying to synchronize to the clock. As a result, stability is not as high as it could be.

Since my ntp time servers based on Intel Pentium processors get their high-speed counter from their processor cycle counter, the idea surfaced that perhaps we could replace the crystal oscillator with a more precise oscillator. This would let the ntp code synchronize a much more accurate timebase, perhaps increasing stability.

So I opened up one of my time servers and had a look at its motherboard. I was quite surprised to see this:

This is a 14.31818Mhz crystal oscillator feeding a PLL that produces the processor's instruction clock. The crystal is clearly visible in the lower left. The PLL is the chip above it.

It seemed like the solution was simply to replace this 14.31818Mhz crystal with a high-precision oscillator of the same frequency. In this case, it's almost too easy.

All I'd have to do is cut the crystal leads carefully and connect the oscillator to one of them. If the change didn't work, I could easily reattach the crystal leads.

But there is one cloud on the horizon. It's not clear whether the PLL will produce a stable output from a stable input. The data sheet for the PLL promises less than 200 ps cycle-to-cycle jitter and .01% frequency stability. But these specifications really give us no clue what will happen.

The conventional wisdom is that it's microscopic temperature variations in the crystal that cause most of the instability seen in ntp time servers. However, it's possible that the instability seen in the synthesized processor clock is actually due to instability in the PLL that generates it from the quartz reference.

I decided I was still going to do it. And as soon as I do, you'll find my results right here.

Back