[Beowulf] AMD performance (was 500GB systems)

[Vincent Diepeveen]

This shows your total ignorence.

At the time several manufacturers claimed specific latencies when  
running random through their SSI setups.

So actually *after* i designed the test... ...which for example  
proved SGI to have hardware factor 12 slower than they claimed to  
have delivered...

A full blown team of guys have been studying the problem as my  
testprogram showed a significant difference in latency depending upon  
the size of the memory block.
Then Paul Hsieh designed a plan and after that it was poured into  
code by another guy who spent a lot of time trying different jump  
patterns through the RAM to measure
whether there was a difference.

They didn't want to see the performance difference between 1 core and  
all cores though - they wanted to make clear the latency
difference between small and larger block and the distance of the  
next read.

Their test never was modified into parallel code though.

If we realize that the manufacturer in this case SGI was factor 12  
off by their latency claim on paper, that's something trivial to you  
obviously.

It had to be proven - it took me time and after that another half  
dozen people a lot of time - yet it's total trivial to you isn't it?

I'm not running with pointer math through the RAM btw, i'm using a  
RNG, so every cpu core is jumping in a different pattern through RAM.
But it seems your lack of understanding didn't grab the difference  
yet when it comes to parallel processing...

On Jan 13, 2013, at 12:46 AM, Mark Hahn wrote:

>> If you call that 'full of yourself', i realize what sort of
>> 'scientist' you are.
>
> Vincent, randomly shuffling a pointer chain is obvious,
> and has been done many times before you did it.