On Mon, Apr 5, 2010 at 12:05 AM, joel jaeggli <joelja@bogus.com> wrote:
On 4/4/2010 7:57 PM, Richard A Steenbergen wrote:
On Mon, Apr 05, 2010 at 10:57:46AM +0930, Mark Smith wrote:
Has anybody considered lobbying the IEEE to do a point to point version of Ethernet to gets rid of addressing fields? Assuming an average 1024 byte packet size, on a 10Gbps link they're wasting 100+ Mbps. 100GE / 1TE starts to make it even more worth doing.
If you're lobbying to have the IEEE do something intelligent to Ethernet why don't you start with a freaking standardization of jumbo frames. The lack of a real standard and any type of negotiation protocol for two devices under different administrative control are all but guaranteeing end to end jumbo frame support will never be practical.
Not that I disagree, given that we use them rather a lot but 7.2usec (at 10Gbe) is sort of a long time to wait before a store and forward arch switch gets down to the task of figuring out what to do with the packet. The problem gets worse if mtu sizes bigger than 9k ever become popular, kind of like being stuck behind an elephant while boarding an elevator.
I didn't run the numbers, but my guesstimate is that would be roughly half the latency that a max sized standard packet would have taken on a 1Gbe switch. It sound reasonable to me that at some point during the march from 10->100->1000->10000 mbit/sec a decision could have been made that one of those upgrades would only decrease max. per hop packet latency by a factor of 2 rather then 10. Particularly since when first introduced, each speed increment was typically used for aggregating a bunch of slower speed links which meant that the actual minimum total latency was already being constrained by the latency on those slower links anyway. OTOH, I totally buy the argument on the difficulty of frame size negotiation and backward compatibility. I think that one of the reasons for the continuing success of "Ethernet" technologies has been implementation simplicity and 100% compatibility above the level of the NIC. Bill Bogstad