> ... I advise to look at your hardware configuration like flow control. > This line doesn't seem good for me (mind the text in bold): > > "Aug 28 15:50:34 ovh-1 kernel: e1000e: enp1s0 NIC Link is Up 100 Mbps Full > Duplex, *Flow Control: None"*
In general, enabling flow control is a bad idea - more so if qos is enabled. I saved this snippet because it's a good explanation of why not to enable flow control (& the snark about the former wife still makes me grin :) >From seif...@netcom.com Tue Mar 23 12:58:14 1999 Newsgroups: comp.dcom.lans.ethernet Subject: Re: 802.3x Flow Control -- Who has it? 802.3x reminds me of my former marriage--at the time, they both seemed like good ideas. However, in retrospect, neither made a lot of sense. =8^) Here's an excerpt on the subject (of .3x, not my former marriage), from my upcoming book, "Make the Switch! A Comprehensive Guide to LAN Switching Technology": --------begin excerpt-------- It is important to remember that the PAUSE function was designed for a very specific use at a very specific time‹to prevent buffer overflow for memory-constrained switches with an input-queued architecture. At the time the protocol was devised (1995-96), many low-cost switches used an input-queued approach, and switch memory was relatively expensive. Reducing the cost of a switch usually meant reducing its memory capacity; by using PAUSE-style flow control, such a switch could provide a good price/performance mix and yet avoid frame loss under peak overload conditions. The PAUSE protocol is less useful today with lower-cost memory and the popularity of output-queued switches. In addition, TCP uses frame loss within the network as its indication of congestion. That is, TCP¹s end-to-end flow control mechanism detects underlying frame loss, and uses this information to throttle its flow. If switches prevent frame loss under congestion conditions, TCP will never recognize the congestion and will continue sourcing data into the network; in fact, TCP will see a lack of frame loss over time as an indication that it can increase the offered load, further exacerbating the congestion problem. If the switches dropped an occasional frame, TCP would throttle back the offered load and alleviate much of the network congestion with no need for PAUSE. This is the concept behind the Random Early Discard (RED) approach to flow control. Implementing link-layer flow control on switches can actually interfere with the end-to-end flow control. --------- Regards, Lee
