click/x-kernel/scout

[Larry Peterson]

    Hi Larry, hope things are going well.
    
Sorry for now getting back sooner. We were frantically trying to meet
yesterday's OSDI deadline. 

    I had some questions about the extent of optimizations in the x-kernel and
    Scout. They may be answered in the papers, but anyway.... This is in
    relation to Click's optimizations. In particular:
    
    (1) How complex are the rules in Scout's transformation rule-based
    optimizer? Can you send us a rules file?
    
Although we have played with other ideas (have you seen our SIGCOMM 96
paper?), generally the way these optimizations get used is to specify
that when a path contains stages a/b/c and attribute foo=bar, replace
the three stages with some stage X. Typically, X is a hand-optimized
stage. For example, we replace stages ETH/IP/ETH with a single
ETH-IP-ETH stage that's optimized for packets that get forwarded
between like devices. In other words, they aren't terribly complex.

    (2) Do the Scout or x-kernel optimizations rewrite C source code -- for
    example, to get rid of function pointers? (In OSDI96, you wrote that paths
    contain composed g(m) functions. Are those g(m) functions compiled with the
    eventual composition in mind? SO for example, if g_1(g_2(m)) is the
    composde function, can g_2 be inlined within g_1?)
    
We have done this, but the current version does not. Hope this helps. --Larry

    Thanks for your time!
    love,
    ed