Subject: Re: FreeBSD Bus DMA
To: None <>
From: David Edelsohn <>
List: tech-kern
Date: 06/12/1998 15:09:00
>>>>> "John S Dyson" writes:

John> I compile exactly the same code, for the same target NetBSD
John> platform, the only difference is running freebsd emulation of
John> NetBSD for compiling the platform.  The compile times compiling
John> EXACTLY the same code, in EXACTLY the same circumstance, on 
John> EXACTLY the same hardware, for EXACTLY the same NetBSD client.

John> We are either cross-compiling the target code under NetBSD, or
John> cross compiling the target code under FreeBSD.  Absolutely
John> no differences in the code being compiled.

>> 1) the code base being compiled
John> Our product.
>> 2) the compile times
John> About 1.5Hrs vs. about 30-40mins or so.
>> 3) the machine types
John> X86. 
>> 4) the versions of the OS that were running.
John> FreeBSD-current, vs. a recent NetBSD.

Separately you also have mentioned 256MB physical memory and that your
entire product including header files, the working set of the toolchain,
and the OS all fit within your physical memory.

	I think that we could have avoided a lot of this discussion if you
had mentioned those details from the beginning.  It also would have been
nice if you had mentioned specific OS releases, compiler options,
filesystem mount options, system tuning parameters, ... as well.

	FreeBSD apparently is self-tuning -- that is a very nice and
useful feature.  NetBSD requires more manual intervention.  I would expect
that NetBSD users who care about performance tune their system -- as most
high performance users do.  Companies who report benchmark results
configure their systems optimally within the limits of the benchmark

	I now can see how you could measure a 2.5X difference in wall
clock time.  It is clear that FreeBSD's caching policy can give it an
advantage for some workloads.  I still do not believe that the ratio you
have quoted corresponds to the actual difference in performance that one
would obtain from optimally configured systems.

	Claiming that a vanilla FreeBSD system can self-tune to better
performance than a vanilla NetBSD system is a different issue.  For users
who care about ease of installation and how much of peak performance they
easily can obtain, this is important.  It is a different measurement than
optimal system performance which is what I, and I presume others, assumed
you were discussing.