UFS, VM, scheduler, emulation questions

From: Orlando Andico <o...@mozcom.com>
Subject: UFS, VM, scheduler, emulation questions
Date: 1999/05/24
Message-ID: <Pine.BSF.3.96.990524015817.2322A-100000@ice.mozcom.com>#1/1
Approved: n...@news1.mpcs.com
Sender: n...@news1.mpcs.com
Content-Type: TEXT/PLAIN; charset=US-ASCII
MIME-Version: 1.0
Newsgroups: mpc.lists.freebsd.questions,muc.lists.freebsd.questions

Hello all,

I'm a new FreeBSD user, although I've used Solaris 1 and 2, IRIX 5, and
Linux for several years and can probably consider myself an experienced
UNIX user. I have several questions about FreeBSD implementations. 

1) FreeBSD 3.x ``feels'' faster than Linux on identical hardware. Is this
a scheduler feature (similar to Windows NT's ``boost foreground
application performance'') or some superior kernel feature that I'm
unaware of? 

2) FreeBSD has a much faster filesystem (according to my simple tests) 
than Solaris 2.6 x86. Of course Linux still holds the record for
filesystem writes because UFS synchronous metadata writes really slow down
filesystem performance. What changes were done to the FreeBSD
implementation of UFS to make this possible? is a UFS+ style metadata
logging (journaling) feature included? 

3) there is an ioctl in SunOS/Solaris to DISABLE synchronous metadata
writes (forgot what it is though..) is this supported on FreeBSD?

4) does FreeBSD support LFS (large file summit) considering that UFS files
can be >>2GB in size according to the FAQ? or is this just PR (e.g. Linux
Reiserfs has 44-bit files, but the libc, etc, don't handle >2GB files) 

5) how are pthreads implemented in FreeBSD 3.x? are these repackaged
Provenzano's threads, ``real'' kernel threads a la Linux (Linux threads
are processes, but fork overhead and ctx switch is very low on Linux as
evidenced by lmbench that it doesn't matter too much), or some other form
of userspace threads (e.g. FSU threads)? 

6) does FreeBSD retain the 4.4BSD/Mach VM system? it certainly consumes
much less memory for buffer cache than Linux, at least the stock setup.
When I run top(1) there is a parameter called "Wired." Is this the same as
wired (nonpageable) memory in the SVR4 VM model?

7) if I want to recompile my kernel and system using PGCC (is this
recommended?) which Makefiles do I alter?

8) can I use DMA on IDE drives like Linux does? what sort of I/O does
FreeBSD do on the wdc device? is it the dumb PIO that Linux does by
default, or is MaxMultsec, etc increased to improve performance?

9) how good is the SMP support in FreeBSD (vis a vis Linux [which is not
that good, even in 2.2] and commercial OS's)?

10) when using memfs on /tmp, is space used for files subtracted from
swap, like in Solaris 2.x? (e.g. if I fill up /tmp with big files, the
system won't be able to swap anymore) 

11) I'm still keeping Linux around to run Informix Dynamic Server (as well
as the ultra-glitzy Red Hat 6.0 GNOME desktop). Will such a complicated
program (IDS) run under emulation? (it uses /etc/shadow instead of
/etc/master.passwd, makes extensive use of glibc2 threads, and does many
select(4096, ..) in there). 

12) how is poll() and select() implemented in the kernel? in the wake of
the Mindcraft fiasco much attention has been given to shortcomings in the
Linux scheduler (particularly in >2 CPU SMP boxes) and the linear scan
necessitated by poll() and select(). How is this issue addressed in BSD?
since wcarchive.cdrom.com handles thousands and thousands of clients, this
issue must have come up sometime.. also, are ``wake one'' select()
semantics implemented? what about sendfile()?

13) (my pet favorite) I like the packaging of Red Hat Linux (it's my other
OS) but the inexplicable ``it feels faster'' performance of FreeBSD makes
me wonder if anyone has done a ``reverse demon Penguin'' -- Linux
userland, FreeBSD kernel. Is this even possible?

14) SGI is releasing XFS under an open-source license. Since XFS is
implemented on top of the vnode/vfs system, conceivably all free UNIX
variants would benefit. Is it worth it to run BSD on XFS, or is UFS
sufficient?

I realize these are a lot of questions, but the FAQ does not cover many of
them. It's easier to find answers on Linux because you can browse the
kernel mailing lists. I was hoping maybe there was a central repository
for questions of this nature.

-------------------------------------------------------------------
Orlando Andico <o...@mozcom.com>                 +63   (2) 937-2293
Mosaic Communications, Inc.                      +63 (912) 800-8262
    Promote bacteria.. it's the only culture some people have.

To Unsubscribe: send mail to majord...@FreeBSD.org
with "unsubscribe freebsd-questions" in the body of the message

From: Oscar Bonilla <oboni...@fisicc-ufm.edu>
Subject: Re: UFS, VM, scheduler, emulation questions
Date: 1999/05/24
Message-ID: <19990524090027.B327@fisicc-ufm.edu>
Approved: n...@news1.mpcs.com
Sender: n...@news1.mpcs.com
References: <Pine.BSF.3.96.990524015817.2322A-100000@ice.mozcom.com>
Content-Type: text/plain; charset=us-ascii
Mime-Version: 1.0
Newsgroups: mpc.lists.freebsd.questions,muc.lists.freebsd.questions

well, since no one has answered this one i might as well givi it a shot.
WARNING: I'm not too sure about the answers. 

On Mon, May 24, 1999 at 02:21:57AM +0800, Orlando Andico wrote:
> 
> Hello all,
> 
> I'm a new FreeBSD user, although I've used Solaris 1 and 2, IRIX 5, and
> Linux for several years and can probably consider myself an experienced
> UNIX user. I have several questions about FreeBSD implementations. 
> 
> 1) FreeBSD 3.x ``feels'' faster than Linux on identical hardware. Is this
> a scheduler feature (similar to Windows NT's ``boost foreground
> application performance'') or some superior kernel feature that I'm
> unaware of? 

i don't think freebsd is just "boosting" the forward application performance
in a multiuser environment there can be many forward applications depending
on how many users the system has. this isn't so in windows nt.

> 
> 2) FreeBSD has a much faster filesystem (according to my simple tests) 
> than Solaris 2.6 x86. Of course Linux still holds the record for
> filesystem writes because UFS synchronous metadata writes really slow down
> filesystem performance. What changes were done to the FreeBSD
> implementation of UFS to make this possible? is a UFS+ style metadata
> logging (journaling) feature included? 
> 
> 3) there is an ioctl in SunOS/Solaris to DISABLE synchronous metadata
> writes (forgot what it is though..) is this supported on FreeBSD?

I think there's a way to disable sync metadata writes... i would check
sysctl(8). Personally I use SoftUpdates.

> 
> 4) does FreeBSD support LFS (large file summit) considering that UFS files
> can be >>2GB in size according to the FAQ? or is this just PR (e.g. Linux
> Reiserfs has 44-bit files, but the libc, etc, don't handle >2GB files) 
> 

hmm, beats me.

> 5) how are pthreads implemented in FreeBSD 3.x? are these repackaged
> Provenzano's threads, ``real'' kernel threads a la Linux (Linux threads
> are processes, but fork overhead and ctx switch is very low on Linux as
> evidenced by lmbench that it doesn't matter too much), or some other form
> of userspace threads (e.g. FSU threads)? 
> 

FreeBSD supports POSIX Kernel Threads (a la linux) and pthreads (user threads).

> 6) does FreeBSD retain the 4.4BSD/Mach VM system? it certainly consumes
> much less memory for buffer cache than Linux, at least the stock setup.
> When I run top(1) there is a parameter called "Wired." Is this the same as
> wired (nonpageable) memory in the SVR4 VM model?
> 

Don't know.

> 7) if I want to recompile my kernel and system using PGCC (is this
> recommended?) which Makefiles do I alter?
> 

No, it's not recomended. FreeBSD 4.0 has egcs as the default compiler and
3.X has gcc 2.7.2.1 as the default. From what I've heard in the mailing lists
compiling the kernel with anything but the default compiler is just a bad
idea.

> 8) can I use DMA on IDE drives like Linux does? what sort of I/O does
> FreeBSD do on the wdc device? is it the dumb PIO that Linux does by
> default, or is MaxMultsec, etc increased to improve performance?
> 

Yes, altough it is not enabled by default. You have to add the flags
0xa0ffa0ff in your kernel config file to each of the IDE controllers.

> 9) how good is the SMP support in FreeBSD (vis a vis Linux [which is not
> that good, even in 2.2] and commercial OS's)?
> 

I don't think any of the opensource OS's have great SMP support.
I would give them fairly good or even usable SMP support.

> 10) when using memfs on /tmp, is space used for files subtracted from
> swap, like in Solaris 2.x? (e.g. if I fill up /tmp with big files, the
> system won't be able to swap anymore) 

I've never used this so don't know.

> 
> 11) I'm still keeping Linux around to run Informix Dynamic Server (as well
> as the ultra-glitzy Red Hat 6.0 GNOME desktop). Will such a complicated
> program (IDS) run under emulation? (it uses /etc/shadow instead of
> /etc/master.passwd, makes extensive use of glibc2 threads, and does many
> select(4096, ..) in there). 

Actually I've just set up informix dynamic server in my laptop under linux
emmulation, i'm accessing it through apache + php3 with the informix module
and since php needs to be linked to the informix libs i had to compile the
whole thing for linux and run everything in linux emmulation mode...

> 
> 12) how is poll() and select() implemented in the kernel? in the wake of
> the Mindcraft fiasco much attention has been given to shortcomings in the
> Linux scheduler (particularly in >2 CPU SMP boxes) and the linear scan
> necessitated by poll() and select(). How is this issue addressed in BSD?
> since wcarchive.cdrom.com handles thousands and thousands of clients, this
> issue must have come up sometime.. also, are ``wake one'' select()
> semantics implemented? what about sendfile()?
> 

Use the source luke :)

> 13) (my pet favorite) I like the packaging of Red Hat Linux (it's my other
> OS) but the inexplicable ``it feels faster'' performance of FreeBSD makes
> me wonder if anyone has done a ``reverse demon Penguin'' -- Linux
> userland, FreeBSD kernel. Is this even possible?
> 

everything's possible if you have source... just how much effort are you
willing to put to accomplish something is a completly different matter.

> 14) SGI is releasing XFS under an open-source license. Since XFS is
> implemented on top of the vnode/vfs system, conceivably all free UNIX
> variants would benefit. Is it worth it to run BSD on XFS, or is UFS
> sufficient?
> 
> 
> I realize these are a lot of questions, but the FAQ does not cover many of
> them. It's easier to find answers on Linux because you can browse the
> kernel mailing lists. I was hoping maybe there was a central repository
> for questions of this nature.
>  

regards,

-Oscar

-- 
For PGP Public Key: finger oboni...@fisicc-ufm.edu


To Unsubscribe: send mail to majord...@FreeBSD.org
with "unsubscribe freebsd-questions" in the body of the message

From: Alfred Perlstein <bri...@rush.net>
Subject: Re: UFS, VM, scheduler, emulation questions
Date: 1999/05/24
Message-ID: <Pine.BSF.3.96.990524132620.9491R-100000@cygnus.rush.net>
Approved: n...@news1.mpcs.com
Sender: n...@news1.mpcs.com
References: <19990524090027.B327@fisicc-ufm.edu>
Content-Type: TEXT/PLAIN; charset=US-ASCII
MIME-Version: 1.0
Newsgroups: mpc.lists.freebsd.questions,muc.lists.freebsd.questions

On Mon, 24 May 1999, Oscar Bonilla wrote:

> well, since no one has answered this one i might as well givi it a shot.
> WARNING: I'm not too sure about the answers. 

You did a good job, i'm just going to tidy up a bit. :)

> 
> On Mon, May 24, 1999 at 02:21:57AM +0800, Orlando Andico wrote:
> > 
> > Hello all,
> > 
> > I'm a new FreeBSD user, although I've used Solaris 1 and 2, IRIX 5, and
> > Linux for several years and can probably consider myself an experienced
> > UNIX user. I have several questions about FreeBSD implementations. 
> > 
> > 1) FreeBSD 3.x ``feels'' faster than Linux on identical hardware. Is this
> > a scheduler feature (similar to Windows NT's ``boost foreground
> > application performance'') or some superior kernel feature that I'm
> > unaware of? 
> 
> i don't think freebsd is just "boosting" the forward application performance
> in a multiuser environment there can be many forward applications depending
> on how many users the system has. this isn't so in windows nt.

I'm not going to pretend I know how it works exactly, but FreeBSD's
scheduler tries to give new and short running processes priority until
they are running for quite some time.

Interactive apps get a bit of a boost, while long running application
are slightly penalized, but only long running applications that are
not interactive.  Don't worry, freebsd won't make your emacs suddenly
get unresponsive.

> > 2) FreeBSD has a much faster filesystem (according to my simple tests) 
> > than Solaris 2.6 x86. Of course Linux still holds the record for
> > filesystem writes because UFS synchronous metadata writes really slow down
> > filesystem performance. What changes were done to the FreeBSD
> > implementation of UFS to make this possible? is a UFS+ style metadata
> > logging (journaling) feature included? 
> > 
> > 3) there is an ioctl in SunOS/Solaris to DISABLE synchronous metadata
> > writes (forgot what it is though..) is this supported on FreeBSD?
> 
> I think there's a way to disable sync metadata writes... i would check
> sysctl(8). Personally I use SoftUpdates.

Going to async metadata is very simple:

mount -o async -u /mountpoint

turning it off:

mount -u /mountpoint

But I agree with oscar, unless you are a specific case that violates the
softupdates license, I would use it:

/usr/src/sys/ufs/ffs/README.softupdates

take a look in there.

> > 4) does FreeBSD support LFS (large file summit) considering that UFS files
> > can be >>2GB in size according to the FAQ? or is this just PR (e.g. Linux
> > Reiserfs has 44-bit files, but the libc, etc, don't handle >2GB files) 
> > 
> 
> hmm, beats me.

I've not had many chances to play with files > 2 gigs except when 
manipulating large tar volumes.  The only problem i've had with large
file is over NFS, otherwise I haven't seen anything weird.

> > 5) how are pthreads implemented in FreeBSD 3.x? are these repackaged
> > Provenzano's threads, ``real'' kernel threads a la Linux (Linux threads
> > are processes, but fork overhead and ctx switch is very low on Linux as
> > evidenced by lmbench that it doesn't matter too much), or some other form
> > of userspace threads (e.g. FSU threads)? 
> > 
> 
> FreeBSD supports POSIX Kernel Threads (a la linux) and 
> pthreads (user threads).

correct, the "linux threads" package is available at:
http://lt.tar.com/

> 
> > 6) does FreeBSD retain the 4.4BSD/Mach VM system? it certainly consumes
> > much less memory for buffer cache than Linux, at least the stock setup.
> > When I run top(1) there is a parameter called "Wired." Is this the same as
> > wired (nonpageable) memory in the SVR4 VM model?
> > 
> 
> Don't know.

Not really, FreeBSD has a unified VM/buffercache but does show some if 
its Mach and 4.4 BSD heritage.  afaik wired does mean non-pageable.

> > 7) if I want to recompile my kernel and system using PGCC (is this
> > recommended?) which Makefiles do I alter?
> > 
> 
> No, it's not recomended. FreeBSD 4.0 has egcs as the default compiler and
> 3.X has gcc 2.7.2.1 as the default. From what I've heard in the mailing lists
> compiling the kernel with anything but the default compiler is just a bad
> idea.

Actually, using pcgcc sounds interesting, you could use it, but you
are entirely on your own here meaning:

1) you fix your makefiles
2) you fix other issues
3) you switch back to the system compiler when unexpected errors
   occur, or fix the problems yourself.

> > 8) can I use DMA on IDE drives like Linux does? what sort of I/O does
> > FreeBSD do on the wdc device? is it the dumb PIO that Linux does by
> > default, or is MaxMultsec, etc increased to improve performance?
> > 
> 
> Yes, altough it is not enabled by default. You have to add the flags
> 0xa0ffa0ff in your kernel config file to each of the IDE controllers.
> 
> > 9) how good is the SMP support in FreeBSD (vis a vis Linux [which is not
> > that good, even in 2.2] and commercial OS's)?
> > 
> 
> I don't think any of the opensource OS's have great SMP support.
> I would give them fairly good or even usable SMP support.

FreeBSD SMP is nice, it works and you get a lot of CPU utilization,
I think the most important portion has been addressed, which is 
stability.  SMP on FreeBSD has been pretty solid for almost the 
last year.  Finer grained SMP is being worked on.

> > 10) when using memfs on /tmp, is space used for files subtracted from
> > swap, like in Solaris 2.x? (e.g. if I fill up /tmp with big files, the
> > system won't be able to swap anymore) 
> 
> I've never used this so don't know.

no, freebsd's /tmp is backed by disk, not swap, however it is entirely
possible to setup what's known as a "MFS" on /tmp, but in most cases
I've been told that simply keepting /tmp as a disk based filesystem
and enabling softupdates make better use of memory and disk bandwidth.

> > 11) I'm still keeping Linux around to run Informix Dynamic Server (as well
> > as the ultra-glitzy Red Hat 6.0 GNOME desktop). Will such a complicated
> > program (IDS) run under emulation? (it uses /etc/shadow instead of
> > /etc/master.passwd, makes extensive use of glibc2 threads, and does many
> > select(4096, ..) in there). 
> 
> Actually I've just set up informix dynamic server in my laptop under linux
> emmulation, i'm accessing it through apache + php3 with the informix module
> and since php needs to be linked to the informix libs i had to compile the
> whole thing for linux and run everything in linux emmulation mode...

ouch, my condolences... :)

> > 12) how is poll() and select() implemented in the kernel? in the wake of
> > the Mindcraft fiasco much attention has been given to shortcomings in the
> > Linux scheduler (particularly in >2 CPU SMP boxes) and the linear scan
> > necessitated by poll() and select(). How is this issue addressed in BSD?
> > since wcarchive.cdrom.com handles thousands and thousands of clients, this
> > issue must have come up sometime.. also, are ``wake one'' select()
> > semantics implemented? what about sendfile()?
> > 
> 
> Use the source luke :)

It's funny how all the ranting and raving didn't make linux faster in
the tests.  

> > 13) (my pet favorite) I like the packaging of Red Hat Linux (it's my other
> > OS) but the inexplicable ``it feels faster'' performance of FreeBSD makes
> > me wonder if anyone has done a ``reverse demon Penguin'' -- Linux
> > userland, FreeBSD kernel. Is this even possible?
> > 
> 
> everything's possible if you have source... just how much effort are you
> willing to put to accomplish something is a completly different matter.

I'm quite sure glibc makes assumption that the Linux kernel is the 
actual kernel, certainly enough hacking and you could probably get
"Linux on a BSD kernel" but, why would you want to?

> > 14) SGI is releasing XFS under an open-source license. Since XFS is
> > implemented on top of the vnode/vfs system, conceivably all free UNIX
> > variants would benefit. Is it worth it to run BSD on XFS, or is UFS
> > sufficient?

Running XFS would be awesome, it's a very cool filesystem.  
The License it's released under will pretty much dictate how FreeBSD
adopts it, whether is makes it into FreeBSD's source tree and where.

> > I realize these are a lot of questions, but the FAQ does not cover many of
> > them. It's easier to find answers on Linux because you can browse the
> > kernel mailing lists. I was hoping maybe there was a central repository
> > for questions of this nature.
> >  

you may search the mailing lists:

http://www.freebsd.org/search/search.html#mailinglists

good luck,
-Alfred 



To Unsubscribe: send mail to majord...@FreeBSD.org
with "unsubscribe freebsd-questions" in the body of the message