once the 100Mb/s link device's buffer is full, you'll end up in a fall-back situation, where you stop sending data to the device until your previous packets are acknowledged (at least, that's how it's supposed to work with TCP. With UDP, the application developer has to handle that themselves). so you'll have very few packets that overflow, if any.My understanding is If it has buffer overflows. TCP/IP packets have to re-transmit if they are dropped. One might see a link at Igb but anytime traffic to the 100m device is involved where it creates a buffer overflow in any port involved, then packets have to be re-transmitted. The general rule is when you are setting up a 1GB network try to keep everything 1GB together, and if you have 100m devices try to keep them downstream on a separate switch or better replace them. I had a 24 port netgear switch at a client's business. One port went to an office in the back to a 10/100 switch. On that switch there were two printers. The person who worked in that office had a separate network cable that came back to another port on the 24 port switch. Three servers also went into the 24 port 1gb switch, The client complained that at times the network was slow. One day a friend of one of the co-owners came in with some high end network analysis equipment that I cannot afford. The first thing he did was replace that 10/100 switch with a 1gb switch. The network problem went away. I know it doesn't make sense but I saw it at the state too so I didn't argue but the slowdowns went away. They transfer some very large autocad files. They also had to upgrade their plotters. Thinking it thru, probably a printer would be less prone to creating a buffer overflow because it isn't typically printing 24/7. A switch would be handling traffic 24/7.
they were probably saturating their uplinks. without QoS enabled, it's really obvious when a (relatively) low priority transfer is stepping on a service that needs higher priority, like VOIP.
in a small environment, I prefer to skip QoS configurations beyond voice (if you even want to bother doing that one), and make sure all uplinks are a bandwidth class above the switchports. so a 48 port gigE switch gets 10G uplinks. now no single user is logically capable of saturating an uplink. my last place had redundant network cores, so each switch got two 10Gig uplinks. each connection would be auto-assigned to the "other" uplink, giving you more bandwidth than having just a failover link.




