Multiple threads efficiency...

Xyzzy · 2004-12-31, 18:53

I frequently run more than one calculation at a time... Basically I run a very low priority GMP-ECM job in the background at "nice -19" and then if I have a quickie I want to run I run it without nice...

It appears that running with those values gives the "nice -19" process 5% of the CPU and the regular process about 90%, with 5% being used by the system...

My question is how much of a loss in efficiency do I get doing this? I'm concerned with cache being swapped in and out all the time and with the cost of context switching... I've played in the past with using a larger quanta but it really messed up interactive performance...

I guess in an ideal situation, GMP-ECM would have a mechanism to pause when another program is running, like Prime95's "PauseWhileRunning" deal...

My big fear is having an idle processor, so by running something in the background all the time I don't have to worry about this... In fact, if I run a job for 100 curves I will frequently start a new job when I get close to finishing just to make sure I don't get any idle time...

Finally, how far can you go? Three concurrent jobs? More?

I guess I could benchmark all of this, but maybe someone already knows the answer and can save me the trouble...

Thanks!

rogue · 2004-12-31, 21:03

Why don't you use ECMNet? It is easy to set up a server and the client can continue to get work for other composites if a factor is found. It is also great if you have a home network and have multiple computers doing ECM factoring. The new ECMNet (v2.5) also supports P+1/P-1 (assuming that you are using GMP-ECM 5.0.3 or above).

2004-12-31, 18:53	#1
Xyzzy "Mike" Aug 2002 79×101 Posts	Multiple threads efficiency... I frequently run more than one calculation at a time... Basically I run a very low priority GMP-ECM job in the background at "nice -19" and then if I have a quickie I want to run I run it without nice... It appears that running with those values gives the "nice -19" process 5% of the CPU and the regular process about 90%, with 5% being used by the system... My question is how much of a loss in efficiency do I get doing this? I'm concerned with cache being swapped in and out all the time and with the cost of context switching... I've played in the past with using a larger quanta but it really messed up interactive performance... I guess in an ideal situation, GMP-ECM would have a mechanism to pause when another program is running, like Prime95's "PauseWhileRunning" deal... My big fear is having an idle processor, so by running something in the background all the time I don't have to worry about this... In fact, if I run a job for 100 curves I will frequently start a new job when I get close to finishing just to make sure I don't get any idle time... Finally, how far can you go? Three concurrent jobs? More? I guess I could benchmark all of this, but maybe someone already knows the answer and can save me the trouble... Thanks!

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Multiple threads per assignment?	f0rteOC	Hardware	3	2016-02-29 06:08
using multiple threads on an LL assignment	tha	Software	4	2016-02-02 13:49
Workers, Threads, Helper Threads, Cores, Affinity.	lorgix	Information & Answers	12	2011-01-13 22:31
pearl question: multiple threads of ECM	Andi47	Programming	2	2010-09-25 06:34
Multiple systems/multiple CPUs. Best configuration?	BillW	Software	1	2003-01-21 20:11

2004-12-31, 21:03	#2
rogue "Mark" Apr 2003 Between here and the 1864₁₆ Posts	Why don't you use ECMNet? It is easy to set up a server and the client can continue to get work for other composites if a factor is found. It is also great if you have a home network and have multiple computers doing ECM factoring. The new ECMNet (v2.5) also supports P+1/P-1 (assuming that you are using GMP-ECM 5.0.3 or above).