FreeBSD & Software RAID

[Wojciech Puchar]

>
>> you talk about performance or if it work at all?
>
> Both, really.  If they have to code up macros to support identical operations

OK. talking about performance:

- 64-bit addition/substraction on 32-bit computer: 2 instructions instead 
of one (ADD+ADC)
- 64-bit NOT, XOR, AND, OR and compare/test etc - 2 instead of one
- multiply - depends of machine, something like 7-8 times longer (4 
multiples+additions) to do 64bitx64bit multiply.
But how often do you multiply 2 longs in C. Actually VERY rarely.

the only exception i can think now is RSA/DSA assymetric key generation 
and processing.

- every operation on 32-bit or smaller values - same
- every branching - same
- external memory access - depends of chipset/CPU not mode - same


now do

cc -O2 -s <some C program>

and look at resulting assembly output to see how much performance could 
really be gained.


about checksumming in ZFS - it could be much faster on 64-bit arch, if 
only memory speed and latency wouldn't be a limit.  and it is, and any 
performance difference in that case would be rather marginal.

> (such as addition) on both platforms, and accidentally forget to use the macro
> in some place, then voila: untested code.
>
>> do you have any other examples of code non-portability between amd64 and
>> i386?
>
> You're also forgetting that this isn't high-level programming where you get to
> lean on a cross-platform libc or similar.  This is literally interfacing with
> the hardware, and there are a whole boatload of subtle incompatibilities when
> handling stuff at that level.

we talked about C code. if not - please be more clear as i don't 
understand what you talking about.

and no - ZFS is not on interface level, doesn't talk directly to hardware.