It is for both. As I said, the switch model has N Interfaces, when Interface::recvPacket is called by system.etherint, it will find the appropriate link connected between the source interface and the destination interface. Yes, I know there are broadcast packets. At least, each system will send arp requests. But I don't think it is a big problem. As described in NSGigE::rxFilter, the switch model can find if the packet is for uni/broad cast. If it is latter, the packet will be sent to all links connected to different system. Otherwise, the switch model will check the destination mac address in packet and find the appropriate link since all mac addresses are known ahead by the switch model. I feel the difference between two switch models is my model doesn't implement the autodiscovery of mac addresses on network.
Xin ---- Original message ---- >Date: Fri, 7 Dec 2007 16:13:16 -0500 (EST) >From: Nathan Binkert <[EMAIL PROTECTED]> >Subject: Re: [m5-users] issues on switch model for cluster >To: M5 users mailing list <m5-users@m5sim.org> > >> My switch model is much like EtherLink model, the bandwidth >> (rate) is a parameter used to create switch model, as it is >> used in EtherLink. >Is the bandwidth managed for the whole switch or for each link or both? >If it's not both, then you're not actually implementing anything >resembling a real switch. If the bandwidth is for just the links, then >you're close, but you'd be doing some sort of infinite rate internally >which isn't what you'd see on a cluster (but it's probably close enough). >If the bandwidth is for the whole switch, then you're just implementing a >hub. Given that you don't even need to implement the EtherLink, why are >you so resistant to using it? > >> I feel my switch model is more easily implemented. >I'd say it's barely easier. If at all. All you need to do if you don't >know a mac address is send the packet out on all ports and then when you >see a packet come in on a port you record the source mac address in an stl >hash_map. You're also not going to succeed if you don't have broadcast >capability because I guarantee that you will see broadcase packets sent >out. > >I guess I was hoping that you'd implement a real switch model and >contribute it so others can use it, but if you're planning on doing it >your way, we can't really accept it without fixing it. > > > Nate >_______________________________________________ >m5-users mailing list >m5-users@m5sim.org >http://m5sim.org/cgi-bin/mailman/listinfo/m5-users _______________________________________________ m5-users mailing list m5-users@m5sim.org http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
