LHC@home

See this WU or that WU for example.

Credit claims from the first three tasks back were 7.55, 11.44, and 25.38 (so the validator correctly awarded 11.44 to everyone, including the tasks that come (will come) back after validation has been done.

I am seeing similar variation quite a lot. Is it my imagination or is has the variation in credit become more extreme recently?

River~~

---

> Hello River~~,
Yes I have also noted this.
I have also noticed that Linux and Windows process the same workunits (I have not been able to find another Linux machine, other than my own, in my last 27 workunits and that is with 4 replications on each work unit).
Depending on who the first 3 in are will depend on what you get, 3 low claimers and you get a low Cobblestone amount, if a couple of slightly higher claiming computers get in early then you can get a few more Cobblestones, luck of the draw.

The other thing I have noticed is that a lot of the work units I get also go to same group of people.
In my first ever batch of wu's, of which I received 11, earlier this month, when I checked the completed jobs the same computer ids kept coming up. Of that 11 wu's that I received 2 others also received the same 11, 1 other got 10 of that 11 and another got 9 of that 11. Meaning that counting me only 8 people shared in the my first batch of wu's, there was 11x5=55 wu's actually issued for this 11 work units.
It improved quite a bit in the second batch I received this week. Of the 17 wu's I received, 10 I shared with one other user, 5 (different wu's) were shared with 3 other users and 4 were also shared with 3 different users. All the rest were spread out over 21 other testers for a total of 29 testers and 17x5=85 work units.
I will have to wait and see if the fairer spread of work continues and this will also change the results that you get. Going on my first batch of wu's with all the same testers I was going to get the about the same cobblestone amount each time.

---

...
Depending on who the first 3 in are will depend on what you get, 3 low claimers and you get a low Cobblestone amount, if a couple of slightly higher claiming computers get in early then you can get a few more Cobblestones, luck of the draw.

absolutely so, that is how it is designed to work.

What I meant in my original post was that the spread of lowest-to-highest cliam within those first three seems to be wider now than it used to be. This is an impression only, I have not kept the figures to back it up.

The other thing I have noticed is that a lot of the work units I get also go to same group of people...

Aha! Good point Conan, and one I had not spotted before.

What is happening here is that a machine gets an allocation of (say) 20 WU. Then another machine later gets an allocation of (say) 17 WU. If the second issue of tasks from the work units is done in the same order as the first issue, then it is likely that the 20 issued to me will overlap with the 17 issued to me, so that maybe 10 of them will be paired with the same machines.

This explains why there are runs of getting low (or high) results out of a single machine -- it is being paired with the same high-claiming (or low-claiming) partner.

It also makes it much more likley that what I think I am seeing here is more of a statistical fluke than genuinely new behaviour.

Thanks for your response, River~~

If you take the time to look at the computers crunching the workunits you'll discover a few things.. in this WU that you mentioned take a look at computer 2207997 and computer 2106539.

What you'll notice is the 1st is a Pentium-M 1.6GHz machine (identical to what I have). The 2nd is a Centrino Core Duo T2300. Dual core CPUs get less credit per workunit here at LHC. There is also a Pentium 4 on that WU with a similar claimed credit, which likely means HyperThreading is enabled on that machine. If the admins don't tweak the credit system here at LHC, dual, quad and so on machine will keep receiving less and less credit per WU. It has to do with the way time is calculated.. 2 cores = double time, 4 cores = quadruple time, etc. That roughly halves (or more) the credit. Rosetta et al have also dealt with this problem, search their boards for some insight.

Credit as far as BOINC is concerned is being generated correctly. The current system doesn't assign credit fairly to more than 1-core processors. Either way the SCIENCE of LHC is (and should be) most important here. When (if) things get back into full swing, they'll most likely address it.

My 2 cents..

---

What you'll notice is the 1st is a Pentium-M 1.6GHz machine (identical to what I have).

Nice chip - the -M is what Intel should have sold instead of the -4, IMO

... It has to do with the way time is calculated.. 2 cores = double time, ... That roughly halves (or more) the credit. ...

I agree totally with your diagnosis, which you explain very clearly.

The main point I was making tho was not about why it is happening, but asking whether this came in with the new servers, or whether people remember the same degree of discrepancy before?

It seems to me that this is a recent effect on LHC, as I don't remember seeing such large variations last time I got work (Easter? May?) and I think I would have noticed. Variations, yes, but not anywhere near as large as this.

You mention tweaks that compensate for these issues: As far as you know, are the tweaks done server side? I am wondering if those tweaks were in there before, and have been lost in the migration to the new server?

If so, given that the incoming admins will be new to all the BOINC stuff, it would be helpful to add a note to the list of things that need looking at - not as a complaint or as a demand, but simply so they know the issue exists to be looked at in their own time.

Or alternatively, is my memory wrong and were we seeing this kind of discrepancy before?

R~~

One of the reasons you might notice this now is that for the first time in several months we're getting WUs again. Several months is a long time for multi-core processors to hit the consumer space and work their way in so there are more of those CPUs here on LHC now. So it wouldn't have been as obvious. This problem isn't new, it's only something that hasn't been observed here for the reasons mentioned until now. No biggie, they can "fix" it (I use the term lightly).

---

The amount of claimed credit also depends upon what version of the boinc client you are running. The mutilple use to determing amount of claimed credit has slowly changed over the versions.

---