Dual nic NAS and Jumbo Frame

Question

I am posting this on the server area because I doubt I am going to get an answer anywhere else.
I have a linux based NAS running netatalk and avahi (afp server and bonjour) with two nics and I have a brand new Mac Pro with two NICS. What I want to do is run a crossover cable between the NAS and the Mac Pro in addition to both being plugged into the normal network. The normal network would have 1500 byte mtu so my internet performance and all of the various vintages of print servers work ok. The dedicated network would have jumbo frames. As we get more Mac Pros, we would add a switch and more machines to this secondary jumbo frame network.
That in theory should work fine (I have done it with other operating systems). My quandary is how to get the Mac to always connect to the NAS via the Jumbo nic and not through the other nic? The Mac learns of the server via Bonjour, so how do I tell it to prefer the "appearance" of the server on the jumbo NIC vs the appearance on the normal network. I know with WINS or DNS I can override the name resolution with a LMHOSTS or hosts file entry, can I do the same with Bonjour?
Thanks for any help or any pointers in the right direction!

Answer 1

I think you are misguided in your assumption that I am not intimately familiar with TCP and don't know what I am talking about.
TCP does not "negotiate" MSS, it advertises the MSS of each side to the remote in the 3 way handshake. It is perfectly acceptable to have asymetric MSS values. TCP does NOT NEGOTIATE a common MSS size. On a LAN, this will result in a functional communication. UDP however does not have such mechanisms and will fail.
TCP will also not function properly in the scnario of my local workstaion with jumbos enabled communicating with a distant endpoint that also has jumbos enabled across a transit network that does not support the maximum MSS used by one of the end stations. For giggles let's say the far end is FDDI and has 4k frame size. Our transit does not support frame sizes larger than the "natural" frame size of 576 bytes. We will use a 4k frame size from me to the remote and a 9k from the remote to me. If the remote sends to me it can use the full 4k MSS of token ring because its less than my MSS. In the reverse my workstation would send 4k frames back to the token ring station. Successful communication would then depend on path MTU and intermediary routers to send ICMP type 3 code 4 messages to signal back to our end stations to reduce our MSS (assuming the DF bit is set on our traffic or the transit router is incapable of fragmentation).
This is perhaps a bit of a flippant example in that nobody would be running FDDI or Token ring anymore, but random entities on the internet will run jumbo frame and perhaps some other l2 technology we aren't familiar with.
Did you ever deal with someone on a token ring segment trying to hit 3Com's web site when it was fddi or token ring? I have on several occasions. I also see this with VPNs all the time. Cisco's genius recomendation is to reduce your MSS on your server as some of their products don't support PMTU. I have had a Cisco <-> Juniper VPN where transfers worked one way because the Juniper would silently strip the DF bit from the packet and fragment it and the Cisco router (38xx) wouldn't do the same in the reverse direction. I also went through **** with the Nortel Contivity VPN devices while they sorted out what to do with the whole MTU negotiation issue.
I have spent many hours of my life pouring through sniffer captures because of mismatched MTUs. Let's not forget the old days of FDDI backbones with ethernet segments bridged across them and FDDI attached servers... mismatched buffers... no thanks.
I therefore don't want to waste my time troubleshooting some bizzare networking issue when there is a perfectly valid way of solving the issue for absolutely minimal expense. I am moving large files here (certainly large enough to get well out of TCP slow start), we easily saturate the full gig link minutes at a time and a saturated gigabit link at standard frame size is inefficient due to the interpacket gap which is locked at 96 bit times for ethernet and the 40 bytes of TCP/IP header plus whatever application payload is prepended per packet on each link. Cutting the number of TCP/IP headers and (probably more importantly since most decent nics do checksum offload these days) application layer headers also reduces load on both client and server.
On large sequential bulk data transfers jumbo frame effectively increases performance and reduces overhead. Period. I have implemented it from the early days of Alteon hardware in Sun servers through Juniper EX products last week. Every iSCSI implementation I run into is jumbo frame based for those exact reasons.
That being said, I don't need to restrict anything. All I want to do is to override bonjour/mDNS for this particular host such that the Pro always communicates over the jumbo segment. This is easily accomplished in windows with an LMHOST entry or in a unix environment with a HOSTS file entry. Is there some way to override bonjour from the client side? I'm ok even statically defining the services presented by bonjour on this host.
I am also willing to force all bonjour requests through a DNS server, however Apple doesn't have any decent documentation on how this is accomplished in an enterprise environment.