Gnu C++ & Open MP

Gnu C++ & Open MP

[burlen]

Hi,
I notice some strange result when I use Gnu OpenMP, I have an existing 
C++ serial code which has some time consuming for loops where some 
calculations are made on a large arrays of doubles,(12 arrays of length 
1E6 for example) and also some of these arrays are copied to create a 
vector field(also an array of doubles). This particular for loop is 
where the code takes 90% of its time and typically can take 1 or 2 
minutes here.  What I notice is that when I run open MP its actually 
slower than the serial code! Not by much but still slower. Here I use 
the command 'time' to compare. Also I notice by watching top, that in 
both cases only 1 core is taxed, while in the OpenMP build memory usage 
doubles(actually a second core goes up from 0% to about 10%). I have a 
quad core system(dual cpu dual core, and 4G of ram), so this result is 
totally unexpected. In my code I explicitly set the number of threads to 
4. I was expecting to see all 4 cores running at 80-90%. I have verified 
that 4 threads are being launched by runing the OpenMP build in GDB. 
What is going on here? how can I figure what the problem is?
Thanks Burlen

Re: Gnu C++ & Open MP

[Tom St Denis]

burlen wrote:
> Hi,
> I notice some strange result when I use Gnu OpenMP, I have an existing 
> C++ serial code which has some time consuming for loops where some 
> calculations are made on a large arrays of doubles,(12 arrays of 
> length 1E6 for example) and also some of these arrays are copied to 
> create a vector field(also an array of doubles). This particular for 
> loop is where the code takes 90% of its time and typically can take 1 
> or 2 minutes here.  What I notice is that when I run open MP its 
> actually slower than the serial code! Not by much but still slower. 
> Here I use the command 'time' to compare. Also I notice by watching 
> top, that in both cases only 1 core is taxed, while in the OpenMP 
> build memory usage doubles(actually a second core goes up from 0% to 
> about 10%). I have a quad core system(dual cpu dual core, and 4G of 
> ram), so this result is totally unexpected. In my code I explicitly 
> set the number of threads to 4. I was expecting to see all 4 cores 
> running at 80-90%. I have verified that 4 threads are being launched 
> by runing the OpenMP build in GDB. What is going on here? how can I 
> figure what the problem is?
> Thanks Burlen

What's the affinity of the threads?  IIRC by default at least with 
pthreads they attach to the CPU that issued the thread so you have to 
set the affinity to 0xF or -1 [all].

Tom

Re: Gnu C++ & Open MP

[burlen]

Tom St Denis wrote:
> burlen wrote:
>> Hi,
>> I notice some strange result when I use Gnu OpenMP, I have an 
>> existing C++ serial code which has some time consuming for loops 
>> where some calculations are made on a large arrays of doubles,(12 
>> arrays of length 1E6 for example) and also some of these arrays are 
>> copied to create a vector field(also an array of doubles). This 
>> particular for loop is where the code takes 90% of its time and 
>> typically can take 1 or 2 minutes here.  What I notice is that when I 
>> run open MP its actually slower than the serial code! Not by much but 
>> still slower. Here I use the command 'time' to compare. Also I notice 
>> by watching top, that in both cases only 1 core is taxed, while in 
>> the OpenMP build memory usage doubles(actually a second core goes up 
>> from 0% to about 10%). I have a quad core system(dual cpu dual core, 
>> and 4G of ram), so this result is totally unexpected. In my code I 
>> explicitly set the number of threads to 4. I was expecting to see all 
>> 4 cores running at 80-90%. I have verified that 4 threads are being 
>> launched by runing the OpenMP build in GDB. What is going on here? 
>> how can I figure what the problem is?
>> Thanks Burlen
>
> What's the affinity of the threads?  IIRC by default at least with 
> pthreads they attach to the CPU that issued the thread so you have to 
> set the affinity to 0xF or -1 [all].
>
> Tom
Thanks for the tip, google'ing I found there is, or rather will be an 
environment variable, GOMP_CPU_AFFINITY, in a later release(its not in 
4.2) I have little desire to configure and build gcc, so I am wondering 
is there an alternate approach? Can I extrenal to OpenMP, control how 
threads are mapped to cpus?

Gnu C++ & Open MP

[burlen]

Hi,
I notice some strange result when I use Gnu OpenMP, I have an existing
C++ serial code which has some time consuming for loops where some
calculations are made on a large arrays of doubles,(12 arrays of length
1E6 for example) and also some of these arrays are copied to create a
vector field(also an array of doubles). This particular for loop is
where the code takes 90% of its time and typically can take 1 or 2
minutes here.  What I notice is that when I run open MP its actually
slower than the serial code! Not by much but still slower. Here I use
the command 'time' to compare. Also I notice by watching top, that in
both cases only 1 core is taxed, while in the OpenMP build memory usage
doubles(actually a second core goes up from 0% to about 10%). I have a
quad core system(dual cpu dual core, and 4G of ram), so this result is
totally unexpected. In my code I explicitly set the number of threads to
4. I was expecting to see all 4 cores running at 80-90%. I have verified
that 4 threads are being launched by runing the OpenMP build in GDB.
What is going on here? how can I figure what the problem is?
Thanks Burlen