Re: how much is my router influencing time-keeping over the network
For fun I'm developing a router for HAM packet networks. What it does is
route AX.25 packets between radios and tunnels (it can also bridge- and
filter them).
What I would like to measure now is, how bad does it influence time-keeping
when syncing time takes place over a network. I could of course just setup
tcp/ip and let two ntp instances sync over it and then calculate an allan
deviation plot.
After an exchange of NTP packets, the client has 4 time stamps.
The time the request packet left the client
The time the request packet arrived at the server
The time the reply packet left the server
The time the reply packet arrived at the client
There are 3 unknowns:
Transit time client to server
Transit time server to client
Clock offset between client and server
With the 4 time stamps, you can setup 2 equations. You need one more.
NTP assumes the transit times are equal.
If you have good clocks at both ends, you can assume the clocks are equal and
compute transit times in each direction.
I would like to see some graphs of network transit times over radio links.
How noisy is yor radio link? It will be interesting to see if if you can get
a decent ADEV graph.
Timings will depend a lot on network traffic. It would be neat if you can get
data under both light load and heavy load. Do you know about bufferbloat? ...
--
These are my opinions. I hate spam.
For fun I'm developing a router for HAM packet networks. What it does is
route AX.25 packets between radios and tunnels (it can also bridge- and
filter them).
What I would like to measure now is, how bad does it influence time-keeping
when syncing time takes place over a network. I could of course just setup
tcp/ip and let two ntp instances sync over it and then calculate an allan
deviation plot.
After an exchange of NTP packets, the client has 4 time stamps.
The time the request packet left the client
The time the request packet arrived at the server
The time the reply packet left the server
The time the reply packet arrived at the client
There are 3 unknowns:
Transit time client to server
Transit time server to client
Clock offset between client and server
With the 4 time stamps, you can setup 2 equations. You need one more.
NTP assumes the transit times are equal.
If you have good clocks at both ends, you can assume the clocks are equal and
compute transit times in each direction.
Indeed, but I also would like measure how good the path between the two
points is. I mean: if the network randomly delays packets (or so), then
that would influence the syncing I think?
In any case, I've done a little experiment with that: I invoke a ping
100k times and then write down the time when the reply was received.
I've explained it further at: https://vanheusden.com/time/ping-test/
I would like to see some graphs of network transit times over radio links.
How noisy is yor radio link? It will be interesting to see if if you can get
a decent ADEV graph.
I'm currently working on that. The two radios are close to each other, with
signal levels of -57dB and SNR of 10.5. Ping times of around 8.3 seconds
(yes, seconds).
Timings will depend a lot on network traffic. It would be neat if you can get
data under both light load and heavy load. Do you know about bufferbloat? ...
I do know about bufferbloat yes.
Hi.
On 2022-09-15 11:02, folkert via time-nuts wrote:
For fun I'm developing a router for HAM packet networks. What it does is
route AX.25 packets between radios and tunnels (it can also bridge- and
filter them).
What I would like to measure now is, how bad does it influence time-keeping
when syncing time takes place over a network. I could of course just setup
tcp/ip and let two ntp instances sync over it and then calculate an allan
deviation plot.
After an exchange of NTP packets, the client has 4 time stamps.
The time the request packet left the client
The time the request packet arrived at the server
The time the reply packet left the server
The time the reply packet arrived at the client
There are 3 unknowns:
Transit time client to server
Transit time server to client
Clock offset between client and server
With the 4 time stamps, you can setup 2 equations. You need one more.
NTP assumes the transit times are equal.
If you have good clocks at both ends, you can assume the clocks are equal and
compute transit times in each direction.
Indeed, but I also would like measure how good the path between the two
points is. I mean: if the network randomly delays packets (or so), then
that would influence the syncing I think?
Yes, it will work on both the mean value and variance. The variance will
increase with additional noise. Some of that will be filtered out by the
control loop, but it will leak through. The mean value of the forward
and backward packet path will vary and not in equal amount, so the
produced time difference will shift along on that difference.
In any case, I've done a little experiment with that: I invoke a ping
100k times and then write down the time when the reply was received.
I've explained it further at: https://vanheusden.com/time/ping-test/
For this I would recommend you to look at MTIE and TDEV rather than
ADEV. ADEV if frequency deviation where as MTIE is a maximum of
systematic in time and TDEV is time deviation.
ADEV and TDEV is intended to handle random noise, and with random noise
we talk about the noises in the Leeson model and not some systematic
effect with what looks like random trafic. I've found that TDEV can be
somewhat of interest too, but MTIE and other TE/TIE types of measures be
much more of interest.
It's only when the time properties is understood that produced frequency
deviation becomes of interest.
I would like to see some graphs of network transit times over radio links.
How noisy is yor radio link? It will be interesting to see if if you can get
a decent ADEV graph.
I'm currently working on that. The two radios are close to each other, with
signal levels of -57dB and SNR of 10.5. Ping times of around 8.3 seconds
(yes, seconds).
Timings will depend a lot on network traffic. It would be neat if you can get
data under both light load and heavy load. Do you know about bufferbloat? ...
I do know about bufferbloat yes.
Measure over a week or two. Compare phase with GPS if you can.
Over a window span, measure average, variance and peak-to-peak for
starters. As buffers fill up and empty, you will for sure see that and
you will see it vary over time if you do measures over suitable
time-periods.
Network noise can be nasty, all dependent on the network and cross-traffic.
Cheers,
Magnus
Since you brought up the speed of light and fiber. One interesting thing
I've noted about white rabbit time sync is the use of BIDI fiber
transceivers. With WR often these are used with one fiber and
bidirectional traffic on it. I assume testing showed this is best using
same path both directions? It's exciting more new network interfaces from
Nvidia and Intel support PTP and making more accurate time stamps. WR is
compatible with IEEE 802.3z Ethernet protocol and IEEE1588v2
https://opg.optica.org/oe/fulltext.cfm?uri=oe-29-8-11693&id=449744
The new Nvidia DPU tech:
https://www.nvidia.com/content/dam/en-zz/Solutions/gtcf21/networking/data-processing-unit/gtc-fall-21-networking-overall-dpu-technical-overview-firefly.pdf
Seems the future is bright with these new DPU's. I'm a network guy but
also a timenut. I wish Cisco included support on more of their new
switches. Hopefully that will change soon. I'm excited today about more
25G connected servers using fan out cables from 100G ports and multigig aka
mGig switches with SFP28 transceivers. Seems to me the time stamping
should be on all ports but alas it's not yet.
Best Regards,
Bill
On Thu, Sep 15, 2022, 3:14 AM folkert via time-nuts <
time-nuts@lists.febo.com> wrote:
For fun I'm developing a router for HAM packet networks. What it does
is
route AX.25 packets between radios and tunnels (it can also bridge- and
filter them).
What I would like to measure now is, how bad does it influence
time-keeping
when syncing time takes place over a network. I could of course just
setup
tcp/ip and let two ntp instances sync over it and then calculate an
allan
deviation plot.
After an exchange of NTP packets, the client has 4 time stamps.
The time the request packet left the client
The time the request packet arrived at the server
The time the reply packet left the server
The time the reply packet arrived at the client
There are 3 unknowns:
Transit time client to server
Transit time server to client
Clock offset between client and server
With the 4 time stamps, you can setup 2 equations. You need one more.
NTP assumes the transit times are equal.
If you have good clocks at both ends, you can assume the clocks are
equal and
compute transit times in each direction.
Indeed, but I also would like measure how good the path between the two
points is. I mean: if the network randomly delays packets (or so), then
that would influence the syncing I think?
In any case, I've done a little experiment with that: I invoke a ping
100k times and then write down the time when the reply was received.
I've explained it further at: https://vanheusden.com/time/ping-test/
I would like to see some graphs of network transit times over radio
links.
How noisy is yor radio link? It will be interesting to see if if you
can get
a decent ADEV graph.
I'm currently working on that. The two radios are close to each other, with
signal levels of -57dB and SNR of 10.5. Ping times of around 8.3 seconds
(yes, seconds).
Timings will depend a lot on network traffic. It would be neat if you
can get
data under both light load and heavy load. Do you know about
bufferbloat? ...
I do know about bufferbloat yes.
time-nuts mailing list -- time-nuts@lists.febo.com
To unsubscribe send an email to time-nuts-leave@lists.febo.com