Path: utzoo!utgpu!jarvis.csri.toronto.edu!clyde.concordia.ca!uunet!cs.utexas.edu!sun-barr!lll-winken!gauss.llnl.gov!casey
From: casey@gauss.llnl.gov (Casey Leedom)
Newsgroups: comp.windows.x
Subject: Re: Remote xterm strategy
Message-ID: <43394@lll-winken.LLNL.GOV>
Date: 7 Jan 90 00:46:07 GMT
References: <90Jan6.054031est.2293@neat.cs.toronto.edu>
Sender: usenet@lll-winken.LLNL.GOV
Reply-To: casey@gauss.llnl.gov (Casey Leedom)
Organization: Lawrence Livermore National Laboratory
Lines: 101


  Mark Moraes explained the trade offs between the two methods well, but
forgot a couple of performance issues.  Expanding slightly on Mark's
diagrams we see some really gross details exposed:

Case 1: rsh remote-host xterm -display local-host:0
---------------------------------------------------
    type:	X -- lh(net) === rh(net) -- xterm -- rh(pty) (*) -- sh/app

    echo:	sh/app -- (*) rh(pty) -- xterm -- rh(net) === lh(net) -- X

Case 2: xterm -e rlogin remote-host
-----------------------------------
    type:	X -- lh(unix) -- xterm -- lh(pty) -- rlogin1
		-- lh(net) === rh(net) -- rlogind2 -- rh(pty) (*) -- sh/app

    echo:	sh/app -- (*) rh(pty) -- rlogind1 -- rh(net) === lh(net)
		-- rlogin2 -- lh(pty) -- xterm -- lh(unix) -- X

Key:
----
    --		User/kernel context switch.
    ===		Network	transfer.
    lh		Local-host kernel.
    rh		Remote-host kernel.
    (pty)	Kernel pseudo tty service.
    (net)	Kernel networking service.
    (unix)	Kernel UNIX domain networking service.
    X		X server running on local host.
    xterm	An xterm client.  Running on the local-host in case 1.
		Running on the remote host in case 2.
    rlogin1	Rlogin process running on local-host.  Reads from local
		xterm via pty interface in CBREAK or RAW mode.  Writes to
		rlogind2 on remote-host.
    rlogin2	Rlogin process running on local-host.  Reads from rlogind1
		on remote-host.  Writes to local xterm via pty interface
		in CBREAK or RAW mode.
    rlogind1	Rlogin daemon process running on remote-host.  Reads from
		sh/app via pty.  Writes to rlogin2 on local-host.
    rlogind2	Rlogin daemon process running on remote host.  Reads from
		rlogin2 on local host.  Writes to sh/app via pty.
    (*)		Sh/app and associated context switch not involved except
		at carriage return if shell/application is in line
		editing mode.

Analysis of case 1:
-------------------
      Typically two context switches on local-host and four on remote-host
    per typed and echoed character when sh/app in line editing mode.
    (Two more context switches per typed and echoed character on
    remote-host when sh/app in non line-editing mode.)  One user process
    executing on each of local-host (X) and remote-host (xterm) in
    typical, line editing mode.  I.e. no user process to user process
    context switches involved.

Analysis of case 2:
-------------------
      Typically ten context switches on local-host and four on remote-host
    per typed and echoed character when sh/app in line editing mode.
    (Two more context switches per typed and echoed character on
    remote-host when sh/app in non line-editing mode.)  Four user
    processes executing on local-host (X, xterm, rlogin1, rlogin2) and
    two on remote-host (rlogind1, rlogind2).  I.e. lots of user process
    to process context switching.

      Also, all characters are forced to go through mainly inactive pty
    code in the local kernel since the local rlogin has put the pty
    interface between itself and the xterm into CBREAK mode (RAW mode
    if the rlogin -8 switch is used).

  Additionally, as Mark mentioned, the larger network packet size for the
X11 protocol versus the rlogin protocol isn't going to make much
difference.  The number of packets transmitted and the inter-packet
spacing are usually the most important features of protocol performance
with respect to Ethernet.  I.e. how much packet room does the protocol
leave?  In this case, the size difference of the X11 and rlogin protocols
is meaningless.  I think there's only one packet pair per typed and
echoed character using an xterm.  If this is so, the total network impact
difference between an rlogin versus an xterm connection is probably
nearly unmeasurable.

Summary and conclusions:
------------------------
  Case 1: executing an xterm on a remote host.

versus

  Case 2: using rlogin to remote-host from within a locally executing
	xterm window.

  Case 2 causes significantly more context switching on the local-host
(ten versus two).  Worse that context switching is among four different
user processes instead of only one.  Even the remote-host which doesn't
see any more context switching is forced to context switch between two
user level processes instead of only one.

  Not having made any actual performance timings it's impossible to make
any conclusions.  But you look at the circuit and tell me that you expect
case 2 to be faster over anything except the slowest networks.

Casey