Re: numerically intensive app doesn't benifit from multi threading
- From: JC <JC@xxxxxxxxxxxxxxxxxxxxxxxxx>
- Date: Mon, 4 Dec 2006 15:53:01 -0800
Here is the info from Dell:
PROCESSOR, 80547, PENTIUM 4 PRESCOTT DT, Pentium 4 Prescott DT, 3.0GHZ, 1
MEGB, 800FSB, SOCKET T, E0, MALE
Can you tell if it's hyberthreading?
Thx.
"JC" wrote:
How can you tell?.
"m" wrote:
Are your 2 CPUs a singe hyperthreaded CPU?
"JC" <JC@xxxxxxxxxxxxxxxxxxxxxxxxx> wrote in message
news:0E9C29A6-F054-4E39-98E3-2FB506F27827@xxxxxxxxxxxxxxxx
This may be related to my another post in this group here "program runs
almost twice as slow on 2 CPU machine when 1 CPU is used".
When I run single thread version of the program on 2 CPU concurently, it's
almost twice as slow as when I run one instance only. When I run multi
thread version of the program with 2 thread on 2 CPU, it's same as running
single thread on 1 CPU.
"JC" wrote:
I have an numerically intensive app where there is a large loop. I want
to
use multithreading to parallelize the loop. Say we have two threads, the
1st
one will do 1st half the loop, the 2nd thread will do the other half.
The
multithreaded loop look like:
for (i = indexBeginForThread; i < indexEndForThread; i++) {
do_numerical_stuff;
update_common_data;
}
The 1st part 'do_numerical_stuff', taking most of the time, and requires
no
mutex to synchronize access to data. The 2nd part 'update_common_data'
takes
little time, and requires mutex to synchronize access to data. You would
expect such a program will have a speed up of 40-50% when running on 2
CPU.
But there is no speend up at all.
I am not sure what is causing no speed up at all.
I inserted 'clock()' function to time the time spend on
'do_numerical_stuff'
and 'update_common_data' in both threads. (I am not sure whether this
works
as I don't know whether clock() is multithread safe). I found that in
both
thead, the total time spent on 'do_numerical_stuff' is the same as when
the
program is run in single thread mode. The total time spent on
'update_common_data' is tiny. So there shouldn't be a problem of waiting
on
the mutex when doing 'update_common_data'.
Theoretically ''do_numerical_stuff' in both thread should use half the
time
as in single thread mode.
The whole loop takes 2 seconds in a test run, and it should benifit from
multi-threading.
Any insights?
Thanks a lot.
- Follow-Ups:
- References:
- Prev by Date: Re: numerically intensive app doesn't benifit from multi threading
- Next by Date: Re: numerically intensive app doesn't benifit from multi threading
- Previous by thread: Re: numerically intensive app doesn't benifit from multi threading
- Next by thread: Re: numerically intensive app doesn't benifit from multi threading
- Index(es):
Relevant Pages
|
