Experiments with PTP
Tue, 13 Dec 2011 11:00:00 GMT
If you ask an historian they would say 1588 was about bad timing; if the Spanish Armada had departed Spain in June instead of late August it might not have been caught by the gales that blew it off course. To the internet community 1588 is the RFC that deals with network synchronisation and the irony is that RFC 1588 is all about accuracy.
There’s been a buzz in the trade tech space about PTP that I’ve been prompted to investigate the new SolarFlare PTP enabled network card. That assumes I can get some time between the ISV test cycles on the low latency infrastructure OnX has built to enable performance tests. The new card has a high quality oscillator that is good enough to act as a grand master clock. The other cards then synchronise on the master or, if there’s a reliable external time source such as GPS or the NPL time signal they could synch to that.
Why is precision time synchronisation so important?
- Test and measurement of different servers in a low latency stack requires more accurate timestamps than is available with standard network time protocol (ntp)
- Packet capture systems without precision time synchronisation requires specialist, expensive network taps that introduce instabilities
- In a regulated industry accurately time stamping transactions provides evidence for dispute resolution.
I’ve been running a series of tests using an FPGA based wire packet capture card that cost $20,000. I would prefer to use the capture capability at the switch and compare application level timestamps on the servers. If I believed the servers and switch clocks were synchronised to within half a microsecond I could capture the data I need without the need for the specialist device. Stay tuned for the results.
David, I believe 1588 is getting a lot of buzz because it's the new thing but NTP is not a stationary target (at least where level of accuracy is concerned). I'm at FSMLabs - we sell TimeKeeper, a software time sync server/client and we show that with both NTP and PTP about the same level of accuracy without hardware assistance. Each give us low microsecond range of accuracy and even better in some setups. With hardware assist as in the solarflare card we see far better - again with both protocols. We've done tests on switches (running our software on them), server, software grandmasters and across the board the protocol choice doesn't seem to affect the accuracy greatly. I think NTP as a protocol has a bad reputation simply because there is a low-accuracy implementation out there that is the golden standard for time on the internet right now.
Cort Dougan 435 days ago