On 2/3/2012 11:00 PM, Andy Bradford wrote: > Thus said Tod Hansmann on Fri, 03 Feb 2012 22:18:47 MST: > >> It's a Netgear GS108t-200NAS, if I recall correctly (I'm not currently >> at the office). I'm actually not too worried about the flooding of the >> switch, as we're only sending 8 MB/s max from the server. Sometimes >> far less (3 MB/s) and we get pause frames not based on the server's >> state, but the clients' states. > It might only be 8MB/s, but what is the size of your packets? If they > are 512 byte UDP datagrams, then that's between 6,144 and 16,384 packets > per second. Are you sure that Netgear can handle that? If the packet > size is even smaller, then increase the PPS and ask again. How big is > the buffer for forwarding packets on the switch? The specs don't seem to > give these details, but I'm not surprised about that for a $100 switch. > What about the rest of the infrastructure? I've never been impressed > with Netgear. They're cheap and poor performers. I wouldn't even > recommend them in a home, and definitely not in a business. > > But, it's also interesting your point about the clients sending the > pause frames. Are they not able to handle the packets? If a client sends > pause frames, that could possibly impact all clients, because there is > only one server, and ethernet frames will slow for *all* clients, not > just the one that sent it, if I'm not mistaken. > Mmmm, you bring up some interesting pieces of info. Two points. The size of the packets is about 1400 bytes, but never varies enough to push it to two frames being needed. I can't remember the exact number of bytes off the top of my head, but it has been brought up as a possibility.
Pause frames may or may not be originating from the clients. We wouldn't know, because when pause frames start occurring at the server, the client is always in a state that we can't check it. It's hung or unconnectable at the very least. However, pause frames are only valid on the link, so they never get propagated through the switch. The server is seeing pause frames from the distributions switch. The switches could be sending pause frames to each other, and clients could be sending pause frames to their switches. We don't know, and finding out would be interesting, but incredibly difficult and expensive to setup. We actually don't mind if packets get dropped. The scenario I described was to sanity check the idea of using an even dumber switch as the distribution switch that doesn't support pause frames to eliminate them from the entirety of the LAN. If we are going too fast, we can account for that and slow down. A nice article on some scenarios where Ethernet flow control causes headaches you don't want (like mine) is here for those interested: http://virtualthreads.blogspot.com/2006/02/beware-ethernet-flow-control.html Cheers, -Tod Hansmann /* PLUG: http://plug.org, #utah on irc.freenode.net Unsubscribe: http://plug.org/mailman/options/plug Don't fear the penguin. */
