[LBo] Performance of Multiprocessor Systems

[Allen]

Sorry, meant SMP, not SNMP which I've been slaving on at work. 
Brain stripped a gear.

Allen

Allen wrote:
> This is very odd seen one way. If 2 cores/CPUs is only 1.5 as fast as a 
> single core due to SNMP overhead, this would imply that a 4 core/CPU 
> machine might not even reach 2x a single core/CPU. If the SNMP overhead 
> is so high then how do super computers with thousands of cores/CPUs get 
> up to teraflops?
> 
> I don't think this is accurate from some points of view as a company I 
> know, Bivio, uses 7 dual core PPC chips and gets almost exactly 13 times 
> the throughput on dnetc.org calcs as I get on a single core 3 GHz Intel.
> 
> Also, my next door neighbor has a dual core Intel at 2.8 GHz running 
> dnetc client and get 13+ million calcs a second and I have a single core 
> 3.0 Intel single core getting 7.1 million calcs per second.
> 
> Allen
> 
> Stefan Waidele wrote:
>> Hi everybody,
>>
>> Andrew Henry schrieb:
>>> mars mr wrote:
>>>> AND NO, if you have Dual CORE at 1,66 GHz you wont
>>>> have same power as 1 processor at 3,2 GHz.
>>>> a second processor at the same speed give around 50%
>>>> more processing power since it has to sincronize with
>>>> the first processor.
>>>>   
>>> Where did you get this information from?  I do not think it is correct
>>> but would like to read your source.  As far as I know, a dual core
>>> processor package runs at the stated GHz speed (1.66 GHz in this case)
>>> but there are 2 distinct cores within the package that split the
>>> workload transparently for the applications
>>
>> According to "Amdahl's Law" 
>> (http://en.wikipedia.org/wiki/Amdahl%27s_Law) the performance of 
>> SMP-systems does not scale ideally with the number of processors.
>>
>> The more processors there are, the more "bookkeeping" needs to be done 
>> to coordinate the processors.
>>
>> While each "core" might run a different (set of) program(s), access to 
>> RAM and cache memory needs to be serialized. The same for harddisk or 
>> network access.
>>
>> Doubling the number of cpus will not double the performance.
>>
>>> --apps do not need to be
>>> multicore aware to be able to use the chips features (this is still
>>> true, despite what I write about processes and multi threading below). 
>>
>> Yes. Even if we ran only one application, the second core could handle 
>> the drawinf of the windows and stuff. Today's systems are so 
>> multi-tasking oriented that a second core will not be idle for long :)
>>
>>> You wrongly refer to core #2 as a second processor and imply that core
>>> #2 runs at 0.5x1.66GHz
>>
>> No, he does not. He assumes the speed increase will be about 1.5 
>> (Twice the raw power minus the overhead in SMP-systems)
>>
>>> [Intel GHz vs. AMD Speedrating]
>>> [...]
>>> If you notice, server chips that are the
>>> best performers only run at 1.8GHz (Xeon/Opteron) whereas desktop
>>> processors are hitting 4GHz, but the server processors are obviously
>>> faster.
>>
>> That shows nicely that "GHz ain't everything".
>> CPU speed also depends on such things as cache-size and strategy, 
>> branch-prediction, the number and granularity of the steps in the 
>> pipeline,...
>>
>> But I want a dual-core, anyways :)
>>
>> Stefan