Experimental results
This page shows the performance of the WRS Low jitter Daughterboard
mounted on a WR Switch (PCB v3.4), compared against the current
performance of an unmodified WR Switch.
The measurements show that the chain Frequency_StandardGM>10km
of fiber->Boundary(Slave) has a stability better than any
commercially available Cesium clocks (ADEV <7E-13 at Tau=1s ENBW
0.5Hz) starting from an integration time as low as Tau=0.1s. The
improvement enable to use White Rabbit to replace a local Cesium
clock.
Telecom engineers, which are more interested on Maximum Time Interval
Error (MTIE) and TDEV, can also benefit from the new stability
performance. The Boundary switch has an MTIE less than 20ps with
Tau=1000s (6h of measurement time) and
TDEV less than 1ps from Tau=0.01s to 1000s*.
Measurement setup
- 2 WR Switch mounting the WRS Low itter board (labelled WRS-HD1 and WRS-HD2) configured as GM and Boundary(Slave)
- 2 WR Switch standard (unmodified PCB v3.4) configured as GM and Boundary(Slave)
- 10 km of fiber spool
- Symmetricom Cs4000 as time reference (frequency standard)
- Microsemi 3120A as phase noise analyzer and stability analyzer (measuring additive phase noise and additive stability)
- Ambient condition: laboratory, temperature fluctuations (not controlled) equals or above +/- 0.5C, no human presence, no direct airflows, HVAC running
- Warm-up time: GM reach full thermal stability after 24h of continuous running. If interested only on stability below Tau=1000s, a warm-up time of 2h is recommended. If interested on stability below Tau=100s, warm-up time of 1h is recommended.
The GM 10 MHz input connector shield is directly connected to the earth of the WR Switch. Be sure that the frequency standard earth connection has not a big loop respect to the earth connection of the GM Switch (suggested setup: same power line strip)*
The GrandMaster switch (WRS-HD1) is connected to the Slave (WRS-HD2)
using Port 2 of the GM and Port 1 of the slave. A traffic generator is
used to create ~1 Megabyte per second (~10Mbit/s) of random traffic to
emulate a real scenario, where the switches are forwarding traffic (port
10 has been used to attach the traffic generator).
Unfortunately only one 3120A is available. I choose to use always the
Cs4000 as time reference (rather than using the 10MHz output of the GM),
in order to proof the stability of the entire time distribution chain
(from a frequency standard to the endpoint). Some users are instead
interested only in the time transfer capability from the GM (ignoring
the time fluctuations between the frequency standard and the GM) to the
endpoint. The GM is less protected by thermal fluctuations since it's in
open-loop (any thermal fluctuation that increase the propagation delay
of the clock line inside the chip is not compensated). The Slave is
protected since any propagation delay would be corrected by PTP.
Due to the choice of using Cs4000 as reference, the experimental results
on long term stability (Tau > 1000s) might be affected by different
working conditions (temperature fluctuations).
Experimental results on GrandMaster
The gateware running on both the GrandMaster (standard and Low Jitter) are based on 18 ports v5.0 (latest release).
Phase noise comparison.
The blue trace is the Standard switch, purple and green traces are from the WR Low Jitter Switch
Allan Deviation comparison (Measurement bandwidth: Equivalent Noise BW 50Hz)
The blue trace is the Standard switch, purple and green traces are from
the WR Low Jitter Switch. The purple trace is an acquisition of 7h,
green trace is from 12h.
The Allan Deviation is not the best tool to characterize the stability
since the phase noise is WPM near the carrier. Due to a bug on the
definition of the Allan Deviation, the stability at 1s is affected by
the phase noise at greater frequencies (e.g. 30Hz) if noise is WPM,
making the result depending on the measurement
bandwidth.
Modified Allan Deviation comparison (Measurement bandwidth: Equivalent Noise BW 50Hz)
The blue trace is the Standard switch, purple and green traces are from the WR Low Jitter Switch. The purple trace is an acquisition of 7h, green trace is from 12h. This is the best tool to characterize the stability of the switch. At Tau=1s you can notice the flicker PM noise (electronic noise of the FPGA), down to Tau=1000s.
Time Deviation comparison. (Measurement bandwidth: Equivalent Noise BW 50Hz)
The blue trace is the Standard switch, purple and green traces are from the WR Low Jitter Switch
Time error
The blue trace is the Standard switch, purple and green traces are from
the WR Low Jitter Switch
The green trace is taken from a GM during a warming up. The purple trace
from a warmed up GM (from 0s to 17000s the peak-to-peak time error is
less than 10ps. The purple trace has a kick at 19000s due to the HVAC
change of modality (the measurement was taken overnight) from night to
daylight.
Experimental results on Slave
The gateware running on both the slaves (standard and Low Jitter) are
based on 18 ports v5.0 (latest release).
The PPSI software is based on latest release (purple) or previous
release (green).
Slave Phase noise comparison.
The blue trace is the Standard switch, purple and green traces are from the WR Low Jitter Switch
Slave MTIE
The figure below shows the MTIE of the green trace (WR Switch using the WRS Low jitter daughterboard) with a measurement time of 6h. For comparison, the MTIE performance of a standard slave switch is around MTIE 100ps at Tau=300s
Slave Allan Deviation comparison (Measurement bandwidth: Equivalent Noise BW 50Hz)
The blue trace is the Standard switch, purple and green traces are from
the WR Low Jitter Switch. The purple trace is an acquisition of 6h using
the latest release of PPSI, green trace is still 6h using the PPSI
release used in WR Switch v4.2.
The Allan Deviation is not the best tool to characterize the stability
since the phase noise is WPM near the carrier. Due to a bug on the
definition of the Allan Deviation, the stability at 1s is affected by
the phase noise at greater frequencies (e.g. 30Hz) if noise is WPM,
making the result depending on the measurement
bandwidth.
Slave Allan Deviation (Measurement bandwidth: Equivalent Noise BW 0.5Hz)
Allan Deviation of the Slave (purple trace) using ENBW 0.5Hz. The result is a significant improvement in the short term stability due to the measurement BW.
Slave Modified Allan Deviation comparison (Measurement bandiwdth: Equivalent Noise BW 50Hz)
The blue trace is the Standard switch, purple and green traces are from
the WR Low Jitter Switch. The purple trace is an acquisition of 6h using
the latest release of PPSI, green trace is still 6h using the PPSI
release used in WR Switch v4.2.
This is the best tool to characterize the stability of the switch. At
Tau=1s you can notice a jump due to the correction done by PPSi. The
jump seems gone away using the latest release of PPSi (the slave clock
reaches the same stability (6.5E-13 at 1s) without PPSi running.
The slight worse stability at Tau=1000s is due to the different thermal
conditions when the acquisition was started (probably due to the GM
stability).
Slave Time Deviation comparison. (Measurement bandiwdth: Equivalent Noise BW 50Hz)
The blue trace is the Standard switch, purple and green traces are from the WR Low Jitter Switch
Slave Time error
The blue trace is the Standard switch, purple and green traces are from the WR Low Jitter Switch
PPS output performance
Measured using the 1-PPS output of the GM and of the Slave Switch
Jumps by the PPSI servo clock
Some jumps were detected using the v4.2 WR Switch software. These jumps
were not coming from the recovered clock by the SerDes. Killing PPSi
resulted in the absence of these jumps. wr_mon showed a set-point
correction value compatible to the jumps measured with 3120A.
These jumps were affecting the stability at Tau=1s. The source of these
jumps seems from an incorrect computation of the correcting values from
PPSi, however due to lack of time an intensive test has not been done.
With the latest release of PPSi the stability is much better, it matches
the stability without PPSi
running.
Phase noise spurs
Some spurs (visible at 20Hz and above) were detected on the phase noise
of both GM and Slave when the Ethernet Management interface is off.
Turning it on (attaching it to a PC) solved the issue. The effect is not
a coupling to the output CLK2 buffer since a spur on the GM is visible
on the Slave.
The effect of the spurs on time stability is negligible.
The picture below show the phase noise spectrum when both the interfaces
(of GM and Slave) are on or off.
This coupling effect has been detected also on the standard switches,
however the effect is less visible due to the higher phase
noise.