Subject: kern/8321: Native ELF execution is slower than Native A_OUT execution
To: None <>
From: The Grey Wolf <>
List: netbsd-bugs
Date: 09/04/1999 10:10:52
>Number:         8321
>Category:       kern
>Synopsis:       Native ELF execution is slower than Native A_OUT execution
>Confidential:   no
>Severity:       serious
>Priority:       high
>Responsible:    kern-bug-people (Kernel Bug People)
>State:          open
>Class:          sw-bug
>Submitter-Id:   net
>Arrival-Date:   Sat Sep  4 09:35:01 1999
>Originator:     The Grey Wolf

NetBSD: We're so committed to it being free, we won't sell it to you
          even if you ask!
>Release:        Userland: 20Aug1999; kernel: 3Sep1999
	Machine: SPARCstation IPX
	OS: NetBSD/sparc 1.4K of 3 September 1999
	Target: Sun4c
	libraries: from binary snapshot of 20 August 1999
System: NetBSD 1.4K NetBSD 1.4K (STARWOLF) #1: Fri Sep 3 20:14:58 PDT 1999 sparc


    SPARCstation IPX (sun 4c/50), BTW.

    Thanks to the 8/20 snapshot I am now running one system under ELF and
    the other under A_OUT.  The emulation is fantastic -- all my old a.out
    stuff works.

    However, I am kind of dismayed at the speed difference here, and I can
    draw but one of two conclusions: i) the file command has changed drastically
    since 1.4-release; or ii) ELF loading is slower than A_OUT loading.

    starwolf is my a.out system (not yet upgraded) and defender is my ELF
    system (the guinea pig, if you will).

    starwolf 13: time file /etc/passwd
    /etc/passwd: English Text

    real	0m1.081s
    user	0m0.476s
    sys	0m0.225s

    defender 15: time file /etc/passwd
    /etc/passwd: ASCII text

    real	0m6.992s
    user	0m2.063s
    sys	0m1.702s

    Overall, I am seeing higher load-and-execute times on my ELF system
    than on my a.out system, though this is by far the most drastic example
    I've found.

    Can someone explain this, and can it be fixed?

    I ran the tests (a.out vs. ELF) on two identical machines down to
    the point that /usr/share was local to both machines (each had their
    own copy), the file on which I ran the test was /etc/passwd which
    was the same length on both machines.

    Moreover, it's startup time which is the problem, not run-time.
    If I passed many files on the command line, file waited for about five
    or six seconds and then processed the files as if there were no
    other difficulties.

    Most of the execution time differences are nowhere near this severe,
    but in general the ELF times are higher than the a.out times.

    Someone mentioned alignment issues.  Anyone else have a clue?

    A good technically explicit answer from a knowledgeable person would be a

    This problem seems germane to the 'file' program, but it indicates
    either a kernel or a compiler problem.

	$ time file /etc/passwd