By: Aaron Spink (aaronspink.delete@this.earthlink.net), September 25, 2007 1:13 pm
Room: Moderated Discussions
Michael S (already5chosen@yahoo.com) on 9/25/07 wrote:
---------------------------
>Not in GBE switching. I didn't check recently, but 2-3 years ago all popular switches
>were store-and-forward. On the other hand, older 100BaseT switches are mostly pass-through.
>
thats cause enet is pretty much crap. Its crap with a huge huge huge installed base but its still crap. enet switches have to be store-and-forward cause of the significant limitations within the enet physical layer itself. basically they made most of the right design choices for the time, but times change, unfortunately, legacy doesn't.
>Could you explain the difference between the terms pass-through, cut-through and
>wormhole? For me all three sound the same when applied to something like Ethernet or EB switch.
>
pass-through is a general term and can apply to both virtual-cut through and wormhole...
wormholes effectively set a continuous transmission from source to destination with the only architected buffering at the end point. with only basic mechanisms to deal with congestions and conflict intra-network.
credit/debit networks rely on having set credits at each point in the network. so by architecture they are store and forward with each point required to have at least 1 credit for every point it talks to.
Virtual cut-through is a mechanism within a credit/debit network to get around the issues of store and forward by using the credit/debit mechanism, but forwarding the packet before it has completely arrived.
Wormhole networks were dominate in the days when putting a couple KB of buffers was extremely prohibitive. It made mor e sense to just do the credit stuff at the end point and retransmit if there was a conflict. AKA enet. This works fine at low congestion or at low speed vs network latency, but as you push throughputs higher and have more things in flight it has issues. Which is why we're effectively doing everything needed for a credit/debit link layer in all currently enet products and have left enet hubs to rot in tar pits with the fossils. But we are still limited by the ethernet architecture in what we can do. so instead of acting like real switches, all the high performance switches in use effectively act like end points in the network and then forward the packets on to the next node in the network.
Aaron Spink
speaking for myself inc.
---------------------------
>Not in GBE switching. I didn't check recently, but 2-3 years ago all popular switches
>were store-and-forward. On the other hand, older 100BaseT switches are mostly pass-through.
>
thats cause enet is pretty much crap. Its crap with a huge huge huge installed base but its still crap. enet switches have to be store-and-forward cause of the significant limitations within the enet physical layer itself. basically they made most of the right design choices for the time, but times change, unfortunately, legacy doesn't.
>Could you explain the difference between the terms pass-through, cut-through and
>wormhole? For me all three sound the same when applied to something like Ethernet or EB switch.
>
pass-through is a general term and can apply to both virtual-cut through and wormhole...
wormholes effectively set a continuous transmission from source to destination with the only architected buffering at the end point. with only basic mechanisms to deal with congestions and conflict intra-network.
credit/debit networks rely on having set credits at each point in the network. so by architecture they are store and forward with each point required to have at least 1 credit for every point it talks to.
Virtual cut-through is a mechanism within a credit/debit network to get around the issues of store and forward by using the credit/debit mechanism, but forwarding the packet before it has completely arrived.
Wormhole networks were dominate in the days when putting a couple KB of buffers was extremely prohibitive. It made mor e sense to just do the credit stuff at the end point and retransmit if there was a conflict. AKA enet. This works fine at low congestion or at low speed vs network latency, but as you push throughputs higher and have more things in flight it has issues. Which is why we're effectively doing everything needed for a credit/debit link layer in all currently enet products and have left enet hubs to rot in tar pits with the fossils. But we are still limited by the ethernet architecture in what we can do. so instead of acting like real switches, all the high performance switches in use effectively act like end points in the network and then forward the packets on to the next node in the network.
Aaron Spink
speaking for myself inc.
