What's the Hurry
29 Jul 2005 3:00:08 UTC
Topic 189635
(moderation:
I just had a bunch of new WU's downloaded and I see that they are all due on 8/4. I've noticed that the time frame allowed for return of completed units in this project is extremely short. Is it that they want only the superfast machines (I've an AMD Athlon 3000+) or want us to crunch only for this project?
THE MOTHER OF FOOLS IS ALWAYS PREGNANT
What's the Hurry
)
Those gravity waves move fast! ;)
Even a slow machine can complete work for this project if your cache settings are set to a reasonable level. This level might be smaller than what you'd use only for projects with longer turnaround times but it's still easily manageable.
Hey Jack, You'll go into
)
Hey Jack,
You'll go into EDF for a while, probably because your connect to value is "too high". Then you'll not get units for a couple weeks and then the whole dance will start over again...
Gotta love LTD =-)
Try here for more info =-)
RE: Hey Jack, You'll go
)
eberndl, are you saying that this is going on again and again?
I set my "connect to network" to 2 days, after I had cleared my cache just after a couple of Einstein WU's nearly got "lost", and with a ressource share LHC: 50%, Einstein: 30%, Seti: 20%, I thought I could keep my Einstein WU's in trim.
So I lowered the "connect to network" to 1 day, and imidiately after I clicked the "Allow new work" button for Einstein, it downloaded 5 WU's!
So the past 3 day, I have crunched solely Einstein WU's to have them all returned before Monday morning!
Is this a way for Einstein to "steal" ressource share? I mean, 5 WU's with short deadline, there won't be much time for my other projects. Even LHC, which has the highest priority for me, can't get much work done right now!
If this is about to go on and on, I'll set the ressource share even lower for Einstein, so it will get the time, I'll give it!
EDIT: BTW, will it help to lower the "connect to network" to 0.1 day???
RE: RE: Hey Jack, You'll
)
The downside to having the preferences on the server is they take effect after you update the project. Meaning when you "update", BOINC first requests work using the old "connect every" setting, then updates its copy of the new preferences. You should see this in the log, a "requesting xxx seconds work" followed by a scheduler reply, followed by a message indicating its using new preferences.
If you just want to update preferences, you can set "no new work" first, then click the "update" button. And after the preferences are updates, set "allow new work".
Walt
RE: The downside to
)
Thanks Walt. I always update my projects imidiately after I have changed the settings. And I check the log to see what it says.
But I think, I'll try to let my 4.72 work it out itself, at least for now! Except I have to update manually to report the finished WU's, which I hope they have resolved in the next version!
My debts are:
PRJ: EINSTEIN@H STD: 0.000000 LTD: -67224.233456 RSRC: 150 -------------------------------
PRJ: LHC@HOME STD: 11830.218725 LTD: 46338.213545 RSRC: 250 -------------------------------
PRJ: SETI@HOME STD: -9409.028835 LTD: 20886.019911 RSRC: 100 -------------------------------
So I don't think I'll get some Einstein WU's today! ;-) But I have at least finished the WU's I had so I won't miss the deadline!
RE: EDIT: BTW, will it
)
I have some machines which are connected to the internet permanently with a broadband connection.
I have set the "connect to network" parameter to 0.01 days!!!
By having this value I only have 1 WU working on and 1.4 minutes before finishing the WU, a new WU is fetched.
I also use parameter sets HOME, SHOOL and WORK to have different values even the machines are not located at 'SHOOL'.
Udo
RE: eberndl, are you
)
Hey Miss Kitty!!!
What it does is not exactly steal time, but BORROW it. So what will happen is you'll d/l let's say 1 day's worth of Einstein (~24 hours)... at a 30% share, your computer will estimate that it will take about 80 hours to complete. Let's also assume that you have enough other work on your machine to put you into EDF, just for kicks.
So, instead of spending 30% of your time on Einstein, you spend lets say 60%... so you spent twice as much time as you should have on Einstein, and so your LTD went into the negative. So you don't down load any Einstein for a while (let's pretend a week), and then you start down loading again.
Again you down load 24 hours worth of work, and again you go into EDF...
So all in all you end up with 1 week of doing 60% Einstein, and then one week doing 0% over the two weeks however, you spent 30% of your time working for Uncle Al.
And this will repeat. again and again and again. just like you don't expect your CPU to be processing 3 units simultaneously with 20, 30 and 50% concurrent resource shares, in order to make the deadline, the CC ensures that you maintain your work share over the long term.
BUT don't feel obliged to change your resource share. As long as you understand WHY you are in EDF (and that EDF is NOT a bad thing, but a very very GOOD thing) you can keep your resource share at whatever you like. I'm at 50% predictor and 8.3% for my other 6 projects.
Finally, yes 0.1 will reduce the chance of EDF, if you have an always on connection and at least 2 projects, I STRONGLY suggest it.
I have 7 projects (G-d help me!!) and have a 0.5 connection time, and that hasn't screwed me over too much yet... but then most of the time I'm only getting work from 5.. and right now only 4 (I want my predictor units back!!!)
Ok, bored now =-)
RE: RE: eberndl, are you
)
Nicely put eberndl!
When it downloads work for Einstein, the max download at any one time is 2*ci (connect interval). That is work for 2*ci assuming E runs alone.
The added complication is that for Einstein, on most machines the scheduler underestimates the work it sends, so in practice you are likely to get (say) 2.8 * ci of actual crunching. You can work out the relevant factor by divding the typical run time of an E wu by the predicted run time before it has started to run. On my machines E wu take about 1.4 times as long to run as they show immediately after download, so my max download in practice will be 2*1.4*ci = 2.8*ci.
If this amount of work is too much to be done at the defined resource share, then this pattern will repeat at intervals of somewhat over a week. The long term effect, controlled by LTD, is that over several months Einstein will keep to its resource share, not by sensible eating but by alternately bingeing and dieting.
If you do not want to see this effect, you have to choose the resource levels big enough, and/or connect interval small enough, so that Einstein can still crunch work for 2.8 * ci within the 1 week deadline despite being pre-empted by other projects.
If you choose not to give E such a generous share of the resources, you are stuck with this effect, but in the long run it will all turn out fair. Just don't get tempted to reset projects to force Einstein to run, or force another project to run: that prevents the clever coding from doing its full magic.
~~gravywavy
RE: Hey Miss
)
Hello Miss Froggy!
Thanks for your explanation, I didn't quite understand this before. I have never had to deal with this before, so when you have mentioned Long Term Debt, I've just glanced over it and shrug. But this started for me when I raised my "connect to" to 2 days because of a planned outage over at Seti, and then Hell broke loose!
Before then I had my "connect to" on 0.5 - 1 and no problems at all! Never more than 1 - 2 WU's at the time and all came and went as they should! I'll reset my "connect to" imidiately!
Thanks! You go, girl!
And thanks to River, Udo, and Walt for your replies. Eberndl has a very good way to explain things, so I understand it! And this is difficult to understand, also because we never had to deal with these concepts before and when you see your computer do something, you don't understand, it's easy to get confused!
I guess I don't see what you
)
I guess I don't see what you mean by having to update manually on 4.72. I am currently running 4.72 on all 4 of my machines. They all upload the results when they are done, and then report them at a later time.
You will see WU's that say "ready to report" after they are done crunching. Just let the reports go by themselves. They will. I promise.
I am looking at one my hosts right now. It has one SETI and one LHC "Ready to Report". Both have a deadline of 8-14-05 (so no worries about deadline problems). it is 1938 7-31-05 (if I am doing my math right, that should be 0038hrs 8-1-05) I'll update this post when I see those two WU's get reported. I promise to NOT do a manual update.
[edit]It is now 2000 local (0100 UTC) and the SETI WU has been reported, hence, the "Ready to Report" message has cleared, with no intervention from me.
Still waiting on the LHC to report, but that could be a while. My current LHC WU still has an estimated 6 hours crunch time.
[/edit]
Jim