So long as the switch's definiton of "jumbo frame" matches that of the NICs on either end. There are some switches out there where "jumbo frame" is not defined as a 9000 byte MTU.
Weelll, that would imply one can run a bus at 1000/1056 or ~95% utilization. One is doing really well (in broad handwaving terms) getting an I/O bus to 80% utilization, and lower than 80% may not be all that surprising.
And that math above does not include stuff like DMA setup latency, bus turn-around events etc etc. So I'd be very pleasantly surprised to see link-rate from a PCMCIA GigE NIC.
So, what is the CPU util and throughput you are seeing with the PCMCIA cards you've used thusfar?
Ah, thanks for the heads up on that; the particular switch I'm looking at says in its docs it supports 9K MTUs.
Oh, me too, but it's not so far off that it's not worth trying to squeeze out some extra performance.
There is no network yet; I'm giving general advice to somebody building a new SOHO-sized GigE network. I'm primarily a Unix server guy, so I'm kind of winging it in the network department. :-)
From what I saw looking around, it looked like the extra cost to support jumbo frames was quite small for other devices (such as switches and PCI cards), and numerous benchmarks showed a performance and CPU utilization improvement. So, it seems sensible to buy a PCMCIA card with jumbo frame support, as long as one is available and the cost isn't too high. We can experiment afterwards to see if it makes a difference or not. On the other hand, if the cards aren't available or are very expensive, we can certainly live without it.
I didn't have any luck finding these cards searching the Internet, and I was hoping somebody here happened to have one and could tell me the brand and whether it was any good.
No problem, I'm winging it in the PCMCIA department. I don't suppose the PCMCIA card vendors say which chips are in their PCMCIA cards do they? You could compare with PCI/PCIX cards and see if it might be a feature that is likely there but not touted or whatnot. Just a wild guess.
Have you tried emailing any of the PCMCIA GigE card vendors?
Another feature you might look for is "TSO" or "large send" support. It is akin to a "poor man's" Jumbo Frame - the sending TCP stack can give large TCP segments to the NIC, which then resegments them into "real" MSS sized segments and spits them onto the wire. Does not require JF support in the receivers or switches - of course it then only really benefits the sender.
It does presuppose support for checksum offload, which if you _are_ going to use JumboFrame, while not required would be very nice because JumboFrame for a bulk transfer will basically cut the number of packets and so the "per packet" costs, but will leave the per-byte costs in place so things become more dominated by the per-byte costs. CKO addresses per-byte costs.