OT Linux (RE: [rescue] OT: Stuffed Proliant?)

[Joshua D Boyd]

On Sat, Dec 22, 2001 at 02:25:35AM -0500, Patrick Giagnocavo wrote:
> Actually, you could probably get better performance from other
> platforms, but it is easier to solve problems with more RAM, which is
> cheap on x86.  DBMS's luuuuuve more RAM.

Absolutely.  But ram for modern Suns isn't so bad.  Don't the netras use 
standard PC100 ram?  Or is that only the SB100s.
  
> All I know is that many people were very happy that the crypto add-in
> cards got supported in OpenBSD and Linux, etc.  Could also be for
> IPSEC/VPN usage though.

Yeah, but they were happy about those cards even for intel systems.
 
>> Besides, lack of rotate isn't all that bad.  Instead of rotating left by 12 
>> bits, you shift left by 12, shift right by 20, then sum the two shifts.  The
>> two shifts can be done simultaneously on a super-pipelined processor, so you
>> need 3 cycles to do a rotate instead of one. While this seems bad, I find it
> 
> As long as the cache stays full...

Well, of course.
 
>>hard to believe that this is the sole reason for distributed net to be slower
>>on SPARCs than intel boxes.  We loose 2 cylces on a rotate.  So what.  We
>>make it up by doing less loads.
> 
> Maybe loads aren't as expensive when everything is in L2 cache.

No, they aren't as expensive, but they still take a minimum of 1 cycle.  
So, doing fewer loads still might balance it out.

I've heard lots of arguments about why Intel is better for encryption, and the
numbers seem to support it, but I really want to know why before I just give 
in.  It's not like it really should be that hard for someone to make a 
definative write-up on the topic, if they just take the time to look at the ASM
and count CPU cycles.

-- 
Joshua D. Boyd