Subject: Re: x86 Port outperforms SPARC64 port on NetBSD v2.0.2
To: David Brownlee <abs@NetBSD.org>
From: Timo Schoeler <timo.schoeler@riscworks.net>
List: port-sparc64
Date: 10/05/2005 12:46:49
David Brownlee wrote:
> On Wed, 5 Oct 2005, John Nemeth wrote:
>
>> } regarding the last paragraph in special, but it's also important to the
>> } former ones: maybe the compiler is crucial? i.e. build NetBSD for your
>> } SS2 with optimization flags and your SSH connection will significantly
>> } speed up :)
>> }
>
>
> Not on an SS2, or any sun4/sun4c machine. Compiler optimisation
> for sun4m and later can win because they support hardware
> multiply, so replacing your SS{1,1+,2} with an SS{4,5,10,20} would
> really help ssh.
i've seen this on SS4, SS5 and SS20MP -- no experience on SS{1,1+,2},
tho. the multiplier issue makes it sound very reasonable ;)
>> } so with bad compiler output (e.g. from gcc, even more w/o optimization)
>> } compared to a surely good-optimizing gcc for i386, that's not very
>> fair.
>> } try to run the software on Solaris and build it with Sun compilers ;)
>>
>> Hmm, I've done it on Solaris, but using GCC. It is very slow:
>>
>> vtn1% time ssh vifa2 id
>> jnemeth@vifa2's password:
>> uid=139(jnemeth) gid=10(staff) groups=10(staff),0(wheel)
>> 0.0u 0.0s 0:27 0% 0+0k 0+0io 0pf+0w
>>
>> vifa2 is a Sparc IPX running SunOS 4.1.3 (aka Solaris 1.1.1). Note
>> that it only took me a couple of seconds to enter my password. Also,
>> it doesn't have a /dev/random, so it is using prng to generate random
>> numbers. This adds to the speed penalty. I am currently putting
>> together another box to replace it.
>
>
> Continuing my above machine reference for lunchbox rather
> than pizzabox systems moving IPC or IPX to Classic or LX would
> give an ssh boost far in excess of any nominal CPU speed gain :)
>
> Pity really, as I really preferred IPX with its cg6 over the
> Classic back when :)
>