TPU's WCG/BOINC Team

[Arjai]

So has a consensus been reached yet on ncpu? Should I add it to BOINC Config Utility or fuggedaboutit?

My conclusion, here on my 3317u i5 2c/2t, is a 25% increase in <ncpus>, will run as if nothing changed. any more and the boincmgr starts to enlarge it's footprint on CPU usage and begins the slowdown. it is up to your own math per CPU to figure if more <ncpus> is worth it. My Math per this CPU?

25% increase. It has been not quite a 25% increase in PPD. Close, I need to have it run a bit longer for it to be a legitimate avg.

 

[FordGT90Concept]

So ncpu 5 on your CPU resulted in ~25% higher PPD?

 

[xvi]

This is a very early report, but my Mobile C2D 1.8GHz laptop appears to be seeing a spike in PPD, although it's hard to tell if it's still on the rise from switching it over to MCM. If the recent increase is correct, it would appear that we're about 10-12% more performance by running 4 WUs on a 2c2t C2D.

 

[TRWOV]

with so few days it's hard to tell, sudden spikes or falls aren't uncommon. I'd wait for the 28 day average before drawning a conclusion.

 

[Norton]

with so few days it's hard to tell, sudden spikes or falls aren't uncommon. I'd wait for the 28 day average before drawning a conclusion.

Agree 100%... there's too much other stuff going on to see what the net positive effect is or if there is any.

 

[Arjai]

So ncpu 5 on your CPU resulted in ~25% higher PPD?

Like I mentioned, I need to let it run a bit longer to get a true result but...

This is points, not avg. The rise is from the 21st to today, the longest runtime since the switch.

Here you can see the results from going 6 <ncpus> to 5<ncpus> with a few other short, 45 minute, tests watching "Performance Monitor".

Daily Avg.:

A 25% increase, or 5 <ncpus> on my 4 logical core i5 is running well. Much better than 6 was, not a little bit because of the BOINCMGR growing quite a bit, CPU-usage-wise, compared to 5, where the mgr barely budged from the stock usage.

I suggest you watch the performance monitor at stock for a spell, switch up <ncpus>, watch it again for a spell. I think the key to a performance increase in PPD is keeping the BOINCmgr, below 10% usage.

Just my observations, feel free to debunk them, as you wish!!

BTW, this is done at 100% CPU on BOINC settings.

 

[Peter1986C]

So has a consensus been reached yet on ncpu? Should I add it to BOINC Config Utility or fuggedaboutit?

It either slows matters down or seems negligible in diff. So forget about it I suggest. Unless the others keep the spikes high even when looking at month-long avgs

 

[ThE_MaD_ShOt]

Well I can say I haven't seen a rise.

 

[FordGT90Concept]

I think I know what is going on here. If you have BOINC set to use less than 100%, ncpu may cause it go higher than your set limit. If you normally have BOINC set to 60% on a quad core with 5 ncpu, for example:

CPU[0] = 60% + 40% = 100%
CPU[1] = 60% = 60%
CPU[2] = 60% = 60%
CPU[3] = 60% = 60%

5 ncpu = 280 / 400 = 70%
4 ncpu = 240 / 400 = 60%

difference = 17%

ncpu can't make BOINC run better than simply setting it to 70% to match it.

 

[Peter1986C]

I believe that "17%" should be "10%". Apart from that, good explanation.

 

[FordGT90Concept]

70/60*100=16.66 ad infinitum

70% is 16.67% more than 60%

Comparing to each other rather than comparing both, individually, to 100%.

 

[Peter1986C]

I see the mistake I made: I compared those two (60 and 70) percentages directly which I should not have done.

 

[ThE_MaD_ShOt]

I agree. I haven't any positive results from setting a higher ncpu. Actually I have been dropping in points since I have been using it. But i already had my rigs set to 95%.

 

[Arjai]

I think I know what is going on here. If you have BOINC set to use less than 100%, ncpu may cause it go higher than your set limit. If you normally have BOINC set to 60% on a quad core with 5 ncpu, for example:

CPU[0] = 60% + 40% = 100%
CPU[1] = 60% = 60%
CPU[2] = 60% = 60%
CPU[3] = 60% = 60%

5 ncpu = 280 / 400 = 70%
4 ncpu = 240 / 400 = 60%

difference = 17%

ncpu can't make BOINC run better than simply setting it to 70% to match it.

It does not run "better." It slows each WU down, all this is while at 100%. The trade off is that now there are 5, or one more WU, being processed and if the slowdown is worth that extra WU.

At 6 WU's, on this computer, the slowdown was not worth the extra WU's. The increase in process time was in hours for each WU. At 5 WU's the slowdown appears to be worth it, as the process time per WU has only risen a few minutes, less than 15min in nearly all cases.

So, very basically, in an hour and 15 minutes, I have an extra done WU.

I hope that makes sense. Said another way, I am paying a price that equates to under an hour and 15 mins. for an extra completed WU.

It seems worth it, to me. Time will tell if it IS worth it.

Sunday night into Monday, I will have a nice long run going, while not connect to WiFi. I will get a real nice snapshot of the process times. The above theory is more than a bit subjective and sometime Monday I will have the numbers worked out.

Like I mentioned, it seems to be working and increasing my PPD. Considering I have a wholly unreliable schedule of runtime, basically 6-8 hours each night augmented for a few hours before, and/or after work.
So, Monday should provide a little enlightenment, at least locally.

Lab Rat, out!

 

[FordGT90Concept]

So you have it set to use 100%? I can't explain that then. Operating System thread switching should be less efficient than not switching at all. I guess having two run on one core could result in simulating a high priority thread but that doesn't make much sense either. Yup, I don't know.


Edit: Oh, its not a dedicated cruncher? Maybe it was crunching longer for the periods it was set to ncpu 5.

 

[Nordic]

@BUCK NASTY, I have not been getting errors on my machine.

My initial assumptions that this would in work nicely on highly overclocked machines might be wrong. I have actually seen a slight decrease overall the last day or two by only 100-300 ppd I believe. Bucks server going up in temp I think is a prime example this does increase load. So My next hypothesis would be that this would help increase load to the maximum and use every available cpu cycle. This seems true based on bucks increase in temps. In my experience, my 4820k usually never really sat at 100%, it would drop down to as low as 97%, and fluctuate between cores. Since I raised the ncpu number every core/thread has never dropped below 100%. I have boinc set to use 100% of cpu too.

Ncpu option could maximum ppd, but only add less than 5%. Probably even less than 2%.

If anyone still feels like testing I believe 50% is much too high still. I think I would say one extra wu for every 3-4 for fast cpus, and probably less for slower cpus. So bucks has a 48 cpu machine but slow cores, so he might 2-5 more wu's.

So has a consensus been reached yet on ncpu? Should I add it to BOINC Config Utility or fuggedaboutit?

I will personally be using this to get that extra 2% or less, because every % matters to me. If you add it, it would see use but not much. You can judge whether or not you see it worth adding. Another use might be running lots of wu's, which will cause them to take longer, so you get badges faster for those into that.




Next I will try the x64 only option and see what happens! I really don't expect anything, expect possibly having less work to run.

 

[FordGT90Concept]

If you really want to squeeze every cycle out of a computer, I would recommend turning the priority up over using ncpus. By default, WUs run low priority so they don't interfere with other software running. If you use no_priority_change 1, they will run at the same as boincmgr which is most likely normal priority. This will make the computer noticeably less responsive but it should yield a nice PPD gain.

Wouldn't 2% be inside the margin of error?

 

[Nordic]

Wouldn't 2% be inside the margin of error?

Practically yes.

 

[Norton]

Went back to 8 cpu's on the 8350 and the ppd increased so I'll hold off for awhile before trying again.

*wondering if changing to only one project would improve anything on the next attempt?

 

[FordGT90Concept]

I'm just thinking that, if I were to add it, I can see people putting 50 in there on a quad core which would translate to a lot of productivity lost. ncpu is something for people that know what they're doing where BOINC Config Utility is aimed at the every-man: useful configuration options most people aren't aware of. In about 3/4 cases so far, "useful" isn't really applicable. I'm strongly leaning towards leaving it out.

 

[Nordic]

I'm just thinking that, if I were to add it, I can see people putting 50 in there on a quad core which would translate to a lot of productivity lost. ncpu is something for people that know what they're doing where BOINC Config Utility is aimed at the every-man: useful configuration options most people aren't aware of. In about 3/4 cases so far, "useful" isn't really applicable. I'm strongly leaning towards leaving it out.

Sounds good.


I also had an idea, that I just mentioned to Norton a bit ago.

Free-dc goes down often. Its a good tool too. What if we TPU WCG team had our own specific free-dc. It only had information about TPU WCG. It hopefully would have more up time, we could have some specific settings that we wanted such as deciding when updates happen. Daily pie and all that could be customized.
There is a lot of possibility here. I have only thought of the ends, but what about the means? How would we check and use data available from WCG? How would we access, read, or use the data?
We would need some sort of database, and a front end website. I was thinking maybe we could run it off of a free amazon ec2 instance.
I am at the limits of my knowledge and ability here. Could someone with more experience maybe provide some insight on the means?

This might be useful information: https://grid.worldcommunitygrid.org/help/viewTopic.do?shortName=overview#329

 

[xvi]

I'm just thinking that, if I were to add it, I can see people putting 50 in there on a quad core which would translate to a lot of productivity lost. ncpu is something for people that know what they're doing where BOINC Config Utility is aimed at the every-man: useful configuration options most people aren't aware of. In about 3/4 cases so far, "useful" isn't really applicable. I'm strongly leaning towards leaving it out.

So, what you're saying is we should have made a bunch of new accounts on RIT and started talking about how <ncpus>200</ncpus> has increased our PPD by, like, 1000%.

 

[Peter1986C]

I think RIT people are too smart or that.

 

[manofthem]

Now that the Challenge is over, i'm going to wind down a bit. My "borrowed" pcs will die off, and I'll now finally be giving F@H a shot on my rig(s), likely tomorrow. But since the folding team has helped us quite a bit, I feel like I need to give back to them a little bit.

 

[TRWOV]

Free-dc goes down often. Its a good tool too. What if we TPU WCG team had our own specific free-dc. It only had information about TPU WCG. It hopefully would have more up time, we could have some specific settings that we wanted such as deciding when updates happen. Daily pie and all that could be customized. There is a lot of possibility here. I have only thought of the ends, but what about the means

I think that donating to Free-DC would be better, that way everybody wins. Otherwise we would need to setup and maintain a server, something that Bok already does.

Boincstats is a good replacement but I really prefer FDCs layout.