Other Parts Discussed in Thread: DP83867CR
Tool/software:
Hello TI-support-Team,
we have a customer that faces some issues with your TI PHY DP83867 and Cisco-Switches.
TI already confirmed this issue at customer's and also our side, so I am creating this ticket to get attention for this topic and to initiate the search for a solution.
Here is our customer's issue description:
they have problems with the PHY DP83867 on SA7.
The problem is called 'IPG/IFG 8 Byte issue', TI already confirmed the topic.
Here you will find some detailed explanation why we see the CRC-errors (loss of data-packages) on the conga-SA7:
On the module we use our Audio-over-IP (AoIP) - device.
AoIP is made of Low-Latency-network with many small packages (>3kHz, package size about 256B, UDP).
When using a Cisco-switch (e.g. from the Series SG350= we see the CRC-errors on RX.
The Switch uses the IEEE-Standard with 8 Byte in RX and uses - dependent on the number and size of the packages, less then 12 Byte IFG/IPG, also 8 Bytes.
These reduces the latency and increases the throughput with small packages.
Also IPG gets used for the orientation, Cisco describes this in their documentation.
The Cisco-switch is reommended by many manufacturers of AoIP-devices. That's why it is used by many end-users. Devices with a different PHY do not have Problems with 8 Byte IFG/IPG. The IEEE-standard for 1GBE also says that RX 8Byte has to be supported.
The CRC-errors are recognized in the Linux Kernel through the MAC and then the packages are lost. The CRC-check can be done by SW and HW. It has no effects on the number of CRC-errors. The Kernel module of the MAC increases the statistics. You can see this with an easy tool like ethtool.
We have done this test with VTM-register. This affects the error, see TI-forum:
DP83867CS: size-dependent packet loss with minimum IPG Part Number: DP83867CS Other Parts Discussed in Thread: DP83867CR , We'd like to understand the minimum IPG supported by this part. Based on a forum post for the e2e.ti.com |
Unfortunately the error is not gone completely. With the undocumented setting 0x3 (see forum) we get the most least errors, so we get about 1 error / hour. 0x2 leads to more or different errors.
TI already confirmed the issues with IPG/IFG 8 Byte on the used PHY.
Maybe it manages up to 10 or 11 Bytes but on lower values it fails.
Here is a description for the test to reproduce the CRS-errors:
When using the Cisco-switch with SA7 you get ~2% errors, that can be shown with ethtool -S.
With a direct connection or using a different switch (normal, no low latency, not managed), you cannot see the CRC-erorrs, as only 12 Byte IFG/IPG are used.
Setting the SA7 to 100MBt/s you will also not see the errors with a Cisco-switch, as the IPG/IFG is much higher.
SA7 (Server):
iperf3 -s -p 5000
PC (Client):
iperf3 -c IP_ADDRESS -p 5000 -u -b 5.85M -P 16 -l 201 --pacing-timer 125000 -t 60 -S 46
The load in the network corresponds to a 'nomal' AoIP-load 50-70 MBit/s)
With a 60s-test you will see about 90 errors (90 packages get lost).
Can you please have a look at this and let me know when the problem will approximately be solved on your side?
Thank you and Best Regards,
Anja Maier