Question Latency through an ethernet switch

[think2]

Does anyone have an idea of how much delay / latency there is in getting a 1500 byte message through a "typical" ethernet switch on 100 Mb/s cable or on fibre. We want to send a "synch clock" message from one device to another through ethernet switches and we need it to arrive as quickly as possible so that the two devices have a synch clock time that is "almost" the same.. We don't want to use NTP. We will have QOS enabled and set the DSCP field set to EF (expedited forwarding) for the synch clock packet so it will be the highest priority packet on the network. I would like to get an idea of how long it would take a packet to traverse, say, 10 ethernet switches.

 

[Hi-Fi Man]

There is significant latency for something like this; honestly can't say how much but that's something you should benchmark. You would probably would be better off using infiniband for anything latency sensitive though.

Here's some links discussing this,

https://www.ibm.com/support/knowledgecenter/en/SSQPD3_2.6.0/com.ibm.wllm.doc/infinibandthrough.html

http://www.hpcadvisorycouncil.com/pdf/IB_and_10GigE_in_HPC.pdf

 

[JackMDS]

May be this Portable Freeware can Help.

https://www.nirsoft.net/utils/network_latency_view.html

 

[mxnerd]

Googled.

https://stackoverflow.com/questions/8682702/how-to-calculate-packet-time-from-latency-and-bandwidth
https://serverfault.com/questions/276651/network-latency-100mbit-vs-1gbit

BTW, I think a time sync does not require a 1500 bytes data frame.

==

https://www.oreilly.com/library/view/ethernet-switches/9781449367299/ch01.html

Forwarding

Store-and-forward​

Refers to standard 802.1D bridging, in which a packet is completely received on a port and into the port buffer (“store”) before being forwarded.​

​

128 KB on-chip packet buffering​

The total amount of packet buffering available to all ports. The buffering is shared between the ports on an on-demand basis. This is a typical level of buffering for a small, light-duty, five-port switch intended to support client connections in a home office.​

​

Forwarding rate

10 Mbps port: 14,800 packets/sec
100 Mbps port: 148,800 packets/sec
1000 Mbps port: 1,480,000 packets/sec​

Latency (using 1500-byte packets)

10 Mbps: 30 microseconds (max)​

100 Mbps: 6 microseconds (max)​

1000 Mbps: 4 microseconds (max)​

​

 

[think2]

Thanks, that's useful information. I don't understand the latency data though. If the latency is the "internal switching delay imposed by the switch electronics", why does it vary with the connection speed? Also does "switch electronics" include any software processing delay?

You're correct, a synch clock packet will be less than 100 bytes but I mentioned 1500 in case there is a 1500 byte packet just in front of the synch clock packet and it stays in front indefinitely because a switch will probably start sending it just before the synch clock packet has arrived. At a 100 MB/s, 1500 bytes takes 120 microseconds to transmit.

If 30 microseconds of latency is typical for a low cost switch then I will be happy.

 

[mxnerd]

I'm not expert.

It seems to me the latency defined in OReilly book is the store and forward delay. Yet it's defined somewhere else as the time it takes for data to go round trip on the network.

Fond some info on the net that you probably will be interested in:
https://www.trincoll.edu/Academics/MajorsAndMinors/Engineering/Documents/IEEE Standard for Ethernet.pdf
(P.41 & P.42 might be interesting)

There is a RFC 2544 standard to test latency: https://www.google.com/search?q=RFC2544+Latency+Test
https://www.marvell.com/documents/rjx203ukari4r93gntem/
https://tools.ietf.org/html/rfc2544
https://www.juniper.net/documentati...ics/concept/oam-rfc2544-testing-overview.html

==

Hope someone in the forum can explain it better.