[LBo] Performance of Multiprocessor Systems

[Michael Cole]

If you have an Intel i have read somewhere the hyperthreading does take a 
knock out of the performance gains, but if you have applications built for 
the multiprocessor system then the speeds will increase.

A program badly built for multiprocessing may only use a single processor for 
all its work but whilst it is running, other processes could be using the 
other processor. 

Depends on the Kernel and the actual programs written for the kernel.

Some programs will not gain anything really as they have to run everything 
step by step others have multiple streams they can be processing concurrently 
so make more use of the power. 

With the Large labs they really have one or more processors just deciding what 
to perform next and where, but the other computers are waiting for the 
functions to run. This is where Gigabit and a well designed network to 
support these farms helps. As all the computers can be accessing data, on the 
network with little or no slow down. If the data all resides on a single disk 
then you have to wait for that disk, and that process to supply the data. 

All comes down to good design.

Also think of the newer Video Cards they are passed a function and some data 
then process that for the Game, the Game programmer does not have to concern 
him/herself with the details of that screen build as much.

The Main processor only passes the function and stream of data, but does not 
have to compute the exact information, thus the speed increase.

This also explains why when new games and hardware come out they test using 3 
or 4 main games and different applications to see if the Game / application 
really gets a boost or not.


On Monday 19 March 2007 22:07:53 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.
-- 
Regards
Michael Cole