| Author |
Message |
|---|
|
|
|
I am running on four i7-processors, now my question is the following:
Each i7 can run 2 WUs simultaneously but they'll each take roughly twice as long as they would if I let each one process just one...
I was wondering if anybody checked which way was more efficient... |
|
|
|
|
|
Hi i don't Know how far your Computer Knowledge goes...
Im Running 8 WU at the same time without any Problem.
Good cooling is essential those days!
If your Processor gets too hot it may Slow down considerably.
First try to find out how to 'optimize' your System.
"hypertreading" may Be a good Start-off
Best regards
Parabel
|
|
|
|
|
I am running on four i7-processors, now my question is the following:
Each i7 can run 2 WUs simultaneously but they'll each take roughly twice as long as they would if I let each one process just one...
I was wondering if anybody checked which way was more efficient...
This has been discussed in various threads but hard to find. The simple answer is that you have to find out what works for you because it depends on the host (Sandy/Ivy Bridge may be different from Nehalem cpus), GPU, the OS (Windows seems different than Linux), projects that you run (boinc and non-boinc like PrimeGrid's PSA) and the PrimeGrid subproject. |
|
|
mfbabb2 Volunteer tester
Send message
Joined: 10 Oct 08
Posts: 510
ID: 30360
Credit: 15,904,188
RAC: 36,681
                     
|
|
Unless you are running 24/7 dedicated to Boinc (no other programs running, etc.), it probably does not make any difference in the long run.
____________
Murphy (AtP)
 |
|
|
 DoESVolunteer tester
Send message
Joined: 11 Oct 08
Posts: 784
ID: 30382
Credit: 75,064,140
RAC: 168
            
|
I am running on four i7-processors, now my question is the following:
Each i7 can run 2 WUs simultaneously but they'll each take roughly twice as long as they would if I let each one process just one...
I was wondering if anybody checked which way was more efficient...
I am thinking you have 1 x i7 CPU = 4 x dual processors = 8 cores (hyper threading)
I have been away for a while but the advantage was always to have as many cores going as possible provided you had the hardware (MB etc) to fully support multi core CPU's
____________
Member of AtP
Shown here is an Australian native rat (Ratus Kickarsus) |
|
|
|
|
I have been away for a while but the advantage was always to have as many cores going as possible
I don't see the advantage....
If I run 2 WUs per core and they take (roughly) twice as long as a single-per-core would, in summary the total number of completed tasks is the same...
The only upside is that you can run twice as many WU at a given point in time;
the downside however is that it's more likely you wind up being a double-checker of a prime (since the tasks take twice as long...)
|
|
|
DoESVolunteer tester
Send message
Joined: 11 Oct 08
Posts: 784
ID: 30382
Credit: 75,064,140
RAC: 168
|
|
What sort of run times are we talking about and what tasks???
my i7 (12 cores) does PPS LLR (680K range) in 550 to 650 seconds (no overclock)
I have 88 PPS LLR primes to my name -- I was double checker only on 7 ocassions
____________
Member of AtP
Shown here is an Australian native rat (Ratus Kickarsus) |
|
|
|
|
|
If I run 1 task per core PPS LLR WUs (in current range) take about 600s,
however with 2 tasks per core it takes about 1200s.... |
|
|
|
|
If I run 1 task per core PPS LLR WUs (in current range) take about 600s,
however with 2 tasks per core it takes about 1200s....
Many of these people just run Windows so don't worry when things make do sense :-).
I have one i7 under Linux and one under Windows but different speeds. So this sounds about right for Linux with Primegrid for PPS LLR and Sophie. Having fewer WUs at once could reduce system stress as the LLR app runs cpus and memory hot. It also frees up resources for other uses.
|
|
|
Michael GoetzVolunteer moderator
Project administrator
Send message
Joined: 21 Jan 10
Posts: 13787
ID: 53948
Credit: 345,142,938
RAC: 21,797
                             
|
Having fewer WUs at once could reduce system stress as the LLR app runs cpus and memory hot. It also frees up resources for other uses.
It probably won't reduce stress on the CPU, at least as long as we're talking about LLR.
Assuming that LLR is one of those programs that is NOT helped by hyperthreading (I believe this to usually be true), it is likely that running 4 WUs will not put less stress (heat, power consumption) on the CPU than running 8 WUs. The reason is that with an application like this, when running 8 'cores' each hyper-core is only running about half the time. The real cores end up doing about the same amount of work either way, thus having the same stress on them. This could be tested by measuring CPU core temperatures with hyperthreading turned on and turned off. (Merely running 4 WUs when hyperthreading is turned on could cause unpredictable test results.)
In applications that DO benefit from hyperthreading, such as the CPU-based sieves, you'll definitely put more stress on the CPU when running 8 WUs, as each real core is doing more work in the same time.
(This might make it sound like running sieves is more detrimental to your CPU than LLR. That's not the case. A single LLR can essentially keep a single hypercore running at 100% of the real-core capacity. That's why using hyperthreading doesn't help any. With sieves, the hypercore runs below 100% of the capacity of the physical core, so using hypthreading allows the other hypercore to make use of the unused capacity of the real core. Hyperthreading just brings the sieve's utilization up to the level of the LLR.)
As for leaving more resources available to other applications, if the resource you mean is memory, than this is certainly true. Other than that, it shouldn't make any difference to other programs or the user interface since BOINC is supposed to run at a very low priority. It shouldn't interfere with anything else, other than making the environment around the computer somewhat warmer
The notable exception to that is with certain GPU applications which seriously hinder UI response. But that's not really relevant to the hyperthreading question.
____________
My lucky number is 75898524288+1 |
|
|
|
|
Having fewer WUs at once could reduce system stress as the LLR app runs cpus and memory hot. It also frees up resources for other uses.
It probably won't reduce stress on the CPU, at least as long as we're talking about LLR.
Assuming that LLR is one of those programs that is NOT helped by hyperthreading (I believe this to usually be true), it is likely that running 4 WUs will not put less stress (heat, power consumption) on the CPU than running 8 WUs. The reason is that with an application like this, when running 8 'cores' each hyper-core is only running about half the time. The real cores end up doing about the same amount of work either way, thus having the same stress on them. This could be tested by measuring CPU core temperatures with hyperthreading turned on and turned off. (Merely running 4 WUs when hyperthreading is turned on could cause unpredictable test results.)
Nice if it was that simple but it is not. Hyperthreading does increase power consumption (various places including Intel) which is not surprising given extra work involved. So, even if small, that translates to less stress and money 'saved'. |
|
|
|
|
Having fewer WUs at once could reduce system stress as the LLR app runs cpus and memory hot. It also frees up resources for other uses.
It probably won't reduce stress on the CPU, at least as long as we're talking about LLR.
Assuming that LLR is one of those programs that is NOT helped by hyperthreading (I believe this to usually be true), it is likely that running 4 WUs will not put less stress (heat, power consumption) on the CPU than running 8 WUs. The reason is that with an application like this, when running 8 'cores' each hyper-core is only running about half the time. The real cores end up doing about the same amount of work either way, thus having the same stress on them. This could be tested by measuring CPU core temperatures with hyperthreading turned on and turned off. (Merely running 4 WUs when hyperthreading is turned on could cause unpredictable test results.)
Nice if it was that simple but it is not. Hyperthreading does increase power consumption (various places including Intel) which is not surprising given extra work involved. So, even if small, that translates to less stress and money 'saved'.
Having said that I decided to talk is cheap so I explored more with my Win 7 i7 940 host. In short, temps are 10oC lower when running 4 WUs at once compared to a full 8 WUs. Also boinc and Windows does not seem to make best use of the CPU with hyperthreading enabled with less than a full load.
I stopped my GPU for this as well. I got a bunch of WUs but clearly was range of values. I used core temp program to monitor as I just grabbed the first one.
With 4 real and 4 virtual cores, host was running at maximum load with peak temps 85-88oC. Three WUs done in 12 minutes and 5 in 12.45 minutes.
Limiting boinc to only 4 WUs at once (suspending or via use 50% of processors) had peak temps between 77-80oC. One set was 10.5 min and another 8.25 min. I did notice that the load was very low typically under 50% and the cpu freq was way way below it should be. This is may explain some of the differences between Windows and Linux not that I am going to that anytime soon.
Then I went and turned off hyperthreading in the bios with everything else the same. Now the load was at 100% with max temps 74-77oC. Ran two sets so 5 WUs took 6.5 min and 3 WUs took 8.25 min.
|
|
|
Michael GoetzVolunteer moderator
Project administrator
Send message
Joined: 21 Jan 10
Posts: 13787
ID: 53948
Credit: 345,142,938
RAC: 21,797
|
Having said that I decided to talk is cheap so I explored more with my Win 7 i7 940 host. In short, temps are 10oC lower when running 4 WUs at once compared to a full 8 WUs. Also boinc and Windows does not seem to make best use of the CPU with hyperthreading enabled with less than a full load.
You didn't explicitly say whether those WUs were LLRs or Sieves, but judging by the run times they have to be LLR.
Given that, I guess I stand corrected. As you said talk is cheap -- and I was greatly simplifying how hyperthreading works, so it's not surprising that the reality differed from what I was expecting to see.
Thanks for running the tests; it's always good to have hard data.
____________
My lucky number is 75898524288+1 |
|
|
|
|
Having said that I decided to talk is cheap so I explored more with my Win 7 i7 940 host. In short, temps are 10oC lower when running 4 WUs at once compared to a full 8 WUs. Also boinc and Windows does not seem to make best use of the CPU with hyperthreading enabled with less than a full load.
You didn't explicitly say whether those WUs were LLRs or Sieves, but judging by the run times they have to be LLR.
Given that, I guess I stand corrected. As you said talk is cheap -- and I was greatly simplifying how hyperthreading works, so it's not surprising that the reality differed from what I was expecting to see.
Thanks for running the tests; it's always good to have hard data.
Sorry, yes they are LLRs. In my view, thanks to the amazing work of people like Ken, sieves should be left to the GPUs.
It was your suggestion and also gave me insight into the difference (or lack of it) between Windows and Linux (there was some recent posting about that). Anyhow back to finding my 100th prime... |
|
|
