Subject: Re: Compiler timings on varous MVII NetBSDs etc.
To: Kevin P. Neal <>
From: NetBSD Bob <>
List: port-vax
Date: 01/23/2001 13:44:12
> On Tue, Jan 23, 2001 at 11:50:04AM -0500, NetBSD Bob wrote:
> > I am afraid I will have to side with Isildur on this one.
> > 
> > But, there is more too it that just feeping creaturitis.
> > 
> > Consider the bare kernel, as one issue.
> > 
> > On my 4.3BSD MVII VAX, the kernel takes 45 MINUTES to compile using
> > pcc IN ONLY 7 MB of ram, and has a footprint of 273K.
> > 
> > On my NetBSD-1.5 MVII VAX, with 13mb of ram, using the SAME config
> > for the kernel (basically stripped down to the minimal BSD bits),
> > the kernel takes 25 HOURS to compile, and has a footprint of 580K.
> Using the same compiler version?

Pcc on the 4.3 box, and gcc-1.91.66 on the NetBSD box.

Recomiling the identical config for a stripped MVII kernel on
4.3BSD, NetBSD-1.2, NetBSD-1.3, NetBSD-1.4.3, NetBSD-1.5,
the times went from 45 minutes to 3 hours, to 5 hours to
9 hours to 25 hours, all on the same machine with identical
hd's loaded with each OS (except for the 4.3BSD load which
was on a 7M ram tiny MVII crate with an esdi mscp disk).
The native compilers were used on each release.
The kernel sizes were 273K in 4.3BSD, 310K in NetBSD 1.2,
(I forget the 1.3 size, but it was aorund 380K), 470K in
NetBSD-1.4.3, and 585K in NetBSD-1.5.

A lot of extra something is growing in the kernel, for a
stripped down KA630 ffs qe0 only config.

> > Consider functionality, e.g., ethernet throughput.
> > 
> > The code has improved in many areas, for example in the tcpip stacks
> > the code has improved speed by almost 100% relative to 4.3BSD in
> > throughput, on NetBSD-1.4.3.  But, 1.5 has dropped back to only a 50%
> > improvement.
> > 
> And this isn't caused by swapping or some other factor outside of the
> network stack?

Dunno.  How could I test that?  The machine it was talking to was
a VAX M76 running NetBSD-1.5.  It should not be a bottleneck for
the slow MVII crate.  The MVII has 13M of ram.  Is ram a bottleneck?
> > Consider compiler efficiency.
> > 
> > A simple program such as gkermit takes 2:50 to compile in Ultrix 4.2,
> > 3:01 to compile in 4.3BSD, and 7:40 to compile in NetBSD-1.5.  What is
> > wrong with that picture?  The compiler is slowing down, tremendously.
> > Interestingly, binary bloat for static binaries had not been a major
> > problem, e.g., gkermit on NetBSD-1.5 was only about 15% bigger than
> > on NetBSD-1.2.
> Is this the same version of gkermit on 1.5 and 1.2?

Same code from Frank da Cruz's kermit archives run up on each machine
under each version of NetBSD.  The native release compilers were used.

> I'm not surprised to see the compiler slowing down. Compilers are an
> excellent example of time vs space tradeoffs and so forth. If the 
> newer compilers are being written to run on modern systems then perhaps
> the code generator is making better code at the expense of compile-time
> CPU (or something). That isn't bloat.

I can see that, but, why the 36 hour non-compilation of perl on the
MVII, and why the 25 hours (25 hours 35 minutes if I am remembering
correctly) to compile the stripped MVII only kernel?  Is it running
out of ram?  Is it something that needs twiddling in gcc setups?
Has the system priced the low-end low-ram boxes out of the market?

> I'm not denying that NetBSD is bigger than 4.3BSD and perhaps slower, of
> course. I'm just curious what exactly is causing the slowdowns you are
> seeing.

I am very curious.  If you get some time, this weekend, or down the
road, stop by and I will let you run the machines and see for yourself.



> Kevin P. Neal