| Index | Recent Threads | Unanswered Threads | Who's Active | Guidelines | Search |
 |
World Community Grid Forums
Category: Community
Forum: Chat Room
Thread: Project performance tables |
| No member browsing this thread |
|
Thread Status: Active Total posts in this thread: 19
|
|
| Author |
|
|
Hypernova
Master Cruncher Audaces Fortuna Juvat ! Vaud - Switzerland Joined: Dec 16, 2008 Post Count: 1908 Status: Offline Project Badges: 
|
I have displayed two tables with the projects ranked according to two parameters that are the ones you can find in the Statistics per Project on the WCG website.
----------------------------------------The Variation parameter is always calculated against the project that ranks 1 in the corresponding table.  Crunchers will select specific projects to crunch for a lot of personal reasons and for them this post is not really of interest. Same for those for whom points or ranking means nothing. But for those who crunch indifferently all projects or are interested in points and ranking then these two tables show that all projects are not equal. We can see that depending on the criteria selected the projects rank very differently in each table. And we see also that between the rank 1 and rank 10 the variation is very large. In the crunchers world we have big differences in the type of hardware. Type (1): single/double core slow/older machines. These machines produce much more runtime but are low in terms of results returned, simply because they will take much more time to crunch each result. Type (2): multicore/hyperthreaded very fast/last gen. machines. These machines produce a very high number of results, as they need less runtime per result and can crunch many results in parallel. The global ranking at WCG is based on points. The question here is what to crunch for each Type of machine to maximize the number of points produced. What we can infer from these two tables is: For Type (1) the top three projects to crunch for are GFAM, DFSL, CEP2. The variation is max -3.4% which means that they are pretty equal. On the bottom, far behind lies FAAH at -48.29%. For Type (2) the top three are the same in a different order: DFSL, CEP2, GFAM. But this time the variation between them is large up to -22.2%. DFSL is a clear winner here. On the bottom very far behind lies HCC at -68.38%. FAAH does a little better here with -42.73% but still not too good. If we cross the two tables we see that: The best performing project overall is DFSL with rank 1 and 2. The worse performing project is HCC with rank 8 and 10. A small additional comment here for those who want to maximize the number of results (and not the points) HCC is the best as it has an average 3 hrs 18 min. runtime per result. It is the shortest not counting DDDT2 (2 hrs 59 min.), but as you know this project is intermittent and to get WU's is like waiting for the Messiah. P.S. If you look at my signature you see that the most crunched project for me was HCC with 30 years. It is time I switch to DFSL and GFAM   [Edit 2 times, last edit by Hypernova at Dec 2, 2011 11:56:29 AM] |
|||
|
||||
|
Former Member
Cruncher Joined: May 22, 2018 Post Count: 0 Status: Offline |
Here a current quick transpose off the WCG Dashboard chart in current credit per hour, [last] 7 day average, accessible to all 24/7sans interruptions due lock out, usually current within the last 12-24 hours
BETA 19.40↓ Gives a quite different story, on the face. Underlying it's deeply complex why each of the anomalies are what they are... or as one of our finance managers once said: It takes years before you can talk to the numbers! --//-- |
||
|
|
Former Member
Cruncher Joined: May 22, 2018 Post Count: 0 Status: Offline |
What we can infer from these two tables is: ... There's one significant factor that you haven't taken into account. The Points per Hour figures from the WCG Statistics per Project are the averages taken over the lifetime of each project. Slower CPUs tend to gain less Points per Hour, and they tend to be the older machines. Therefore, the longer-running projects will generally show a lower project-lifetime Points per Hour, and conversely the more recent projects will give a higher figure. Your ranking by Points per Hour is almost exactly the same as you would get if you ranked by age of project since each launch date. So I think the main inference from your table is that some projects are older than others! Sek's table of current 7-day credit per hour is a better way to deduce the overall effect of different projects on points, and that shows much less variation (as is no doubt intended). Nevertheless, individual machines & operating systems do show a variation from even that community-wide current average, so anyone keen to maximise some aspect of their processing has to work out their own figures over a week or two. |
||
|
|
Hypernova
Master Cruncher Audaces Fortuna Juvat ! Vaud - Switzerland Joined: Dec 16, 2008 Post Count: 1908 Status: Offline Project Badges:
|
After having looked at the Sekerob stats, I double checked my tables, and I edited my previous post. No changes except that some values were not correct (excel cut and paste typical issues). It does not change anything regarding my analysis based on those values.
----------------------------------------Now here is the Dashboard table with variations.  Indeed a different situation. Still the difference between the two extremes is 17% which is much less but still not negligible. Rankings are very different too. We see that HCC also here ranks 7 so there is some coherence at least for HCC. I agree with tonyh205 comments. Because the values I use are global averages indeed the project duration will have an effect. It would be interesting to measure that effect  |
||
|
|
Former Member
Cruncher Joined: May 22, 2018 Post Count: 0 Status: Offline |
Suggest you remove DDDT2 from the table as that is only a 1 day stat, driven by fast returners [guess what devices these are] and small population. What you then get is...
--//-- |
||
|
|
cjslman
Master Cruncher Mexico Joined: Nov 23, 2004 Post Count: 2082 Status: Offline Project Badges: 
|
Very interesting discussion indeed ! Some questions (because I'm probably in this situation):
----------------------------------------- Are the points granted differently between computers (one older/slower and one newer/faster) for the same WU ? For example, let's suppose that one computer took 12 hrs to crunch the WU and the new one took 8 hrs (assuming that the WU is exactly the same for both). - Does the older computer get more points for taking longer to crunch the WU? - Does the new computer get more points for crunching the WU faster? Thanks, CJSL |
||
|
|
Hypernova
Master Cruncher Audaces Fortuna Juvat ! Vaud - Switzerland Joined: Dec 16, 2008 Post Count: 1908 Status: Offline Project Badges:
|
Here again my performance tables versus project duration.
---------------------------------------- As tonyh205 stated there is indeed an excellent correlation between rankings in Points/Runtime and Project duration. The older the project and "less performant" is the project and vice-versa. But if we look at the Point/Result table there is no evident correlation here. Maybe it is more difficult to find an underlying reason as this is related to the type of machines in the grid crunching for projects. But if we consider that we have a global average in terms of hardware all over the grid then the Point/Result is the best criteria to judge project performance, as it is not directly correlated to project longevity. cjslman to answer your question, the new computer will get more points. If you want to maximise points use as faster as possible CPU with as much possible cores with hyperthreading.  |
||
|
|
kateiacy
Veteran Cruncher USA Joined: Jan 23, 2010 Post Count: 1027 Status: Offline Project Badges:
|
Another factor affecting these numbers is that people would tend to run the smaller, less-demanding sciences on slower machines. I know I run HCC and HCMD2 on my Atoms and run CEP2, HPF2, etc. only on my faster machines.
----------------------------------------That said, I have observed that when I was running HCC on my fast machines to finish a badge, HCC earned fewer points per hr than CEP2, GFAM, and DSFL do on the same machines. Also, my "Performance" on Snurk's signature has climbed by several Gflops since I started running mostly GFAM and DSFL instead of mostly FAAH and HPF2 on my faster machines. We might want to share some results from Pirogue's WCGDaws comparing different sciences running on the same machines.  |
||
|
|
krakatuk
Advanced Cruncher Germany Joined: Oct 3, 2008 Post Count: 141 Status: Offline Project Badges: 
|
Hypernova,
----------------------------------------Point/Result statistics shows us only one thing: some projects have smaller WU's which need less time and bring less points. Another projects have long-running WUs, which bring you more points. From your tables you can see that HCC has the smallest WU size, where CEP2 & DSFL - biggest. So Point/Result is not a good choice to mesure project performance... What I can say after some experience of crunching WCG on some very different boxes - the main variable to find out which project performs better for you - is your hardware. I've found out that my hyperthreaded device tends to claim more points than it usually gets. Non-HT devices claim less. That means for me, that the best project to run on my HT-device is CEP2 (they give as much as you claim now). For non-HT devices the best project to run is HPF2 (they have 19 copies of every WU and you have always some systems out of this 19 who claim more and bring the granted amount of points into the hight). When I crunch only these projects on the right boxes I get about 20% more points. But usually I don't care so much about points and crunch what I currently want. Honestly saying I don't think this difference in points means that those projects perform better on that hardware. Probably it's just HT, which is cheating during a benchmark. Edit: Regarding the cjslman's question - all computers running the same WU get same amount of points for this WU. Usually it's just an average from what they all claim.  [Edit 3 times, last edit by krakatuk at Dec 2, 2011 2:26:58 PM] |
||
|
|
cjslman
Master Cruncher Mexico Joined: Nov 23, 2004 Post Count: 2082 Status: Offline Project Badges:
|
Thanks Krakatuk for answering my question... that is exactly what I would expect: The same WU should give the same amount of points independent of the performance of the CPU that crunched it.
----------------------------------------CJSL |
||
|
|
|