[parsec-users] serial version of parsec suite

Zhunping Zhang zhunping at csail.mit.edu
Thu Feb 2 05:44:18 EST 2012


So there is a communication channel between the processor and the  
memory, and this channel has limited bandwidth. For example, if the  
bandwidth is X, and running one copy of a program P takes up 100% of  
the bandwidth, then running two copies of the program at the same time  
would slow each of them down by 2x, because now they need to share the  
memory bandwidth. In some other case, this is ok, for example, if  
running one copy of the program only takes 10% of the channel, then  
you should in theory be able to run 10 copies of it without seeing  
slowing down due to memory bandwidth.

So check if your program is memory bandwidth intensive, simply run P  
copies of the same program, e.g., ferret. and compare the time to the  
one you got by running only one copy of it. If the times are  
relatively the same, then the program is not memory bandwidth limited.

In general, it is not a good idea to mix multiple programs to get  
numbers. There might be other factors that let program influence each  
other, such as cache, disk. So SPEC got its published numbers by  
running different programs on different cores at the same machine?

Cheers,
Justin


On Feb 2, 2012, at 5:27 AM, Zhunping Zhang wrote:

>> CMP

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.cs.princeton.edu/pipermail/parsec-users/attachments/20120202/dbab2c0a/attachment-0001.html>


More information about the parsec-users mailing list