Re: draft-cameron-tmux-02.txt

["Donald E. Eastlake 3rd (Beast)" <[e-mail]>]

From:  Pete Cameron <[e-mail]>
To:  [e-mail]
Cc:  [e-mail]
>---------  Received message begins Here  ---------
>
>> From: "Donald E. Eastlake 3rd (Beast)" <[e-mail]>
>
>Donald,
>
>Thanks for commenting again.

You're welcome.

>> I've always thought that the tmux concept was very elegant and general
>> and should apply to all IP traffic.  I mean it just groups packets for
>> efficiency.  At a high level, why should it matter a bit what kind
>> of packets they are?
>
>The consensus of opinion at the Houston BOF was that the
>document should be made very specific to terminal servers,
>hence the direction I am taking. Also, it matters a LOT what
>kind of packets are passed from the higher level to TMux. We
>tried some experiments with ftp to see what would happen. If
>we allow ftp packets to be multiplexed, then for a single
>ftp session we got a 15% degradation in performance. We
>consider this unacceptable for something that is designed to
>increase performance. Having said this, I still agree to a
>degree with your sentiments, TMux is applicable to all IP
>trafic, but must be used with care.

Well, why was that?  You already recommend against tmux'ing large
packets, which is reasonable.  Was it the delays in ACKs or something?

Even if this caused degradation for a single ftp session, would it have
increased throughput if there had been 10 ftp sesssions?

It would be excellent if you could characterize what benefits most
from tmux other than saying "terminal servers"...

>> It seems to me that you might want to be able to configure a host to
>> the opposite assumption, i.e., that all other hosts are tmux capable
>> unless you get a ICMP error back in which case you add it to your list
>> of non-tmux capable hosts and time out that list with some long
>> time-out.  This might be appropriate for a terminal server.  Note:
>> this eliminates the need for any ENQ and/or RSP!
>
>Our first attempt at implementing TMux used exactly this
>idea.  Unfortunately the BSD TCP/IP (as used by many hosts)
>does not send ICMP errors for unknown protocols.  This is
>because all unknown IP protocols are passed upto the RAW
>protocol driver where they fall into a black hole :-(

BSD bugs strike again!


>Thanks
>
>Pete Cameron                                            tel: [phone]
>Consulting Engineer                                     fax: [phone]
>Xylogics International Limited                  email: [e-mail]
>Featherstone Rd, Wolverton Mill, MK12 5RD, UK.         [e-mail]

Donald