Welcome to MilkyWay@home

So you want us to try to beat your supercomputer.

Message boards : Number crunching : So you want us to try to beat your supercomputer.
Message board moderation

To post messages, you must log in.

AuthorMessage
Bigred
Avatar

Send message
Joined: 23 Nov 07
Posts: 33
Credit: 300,042,542
RAC: 0
Message 615 - Posted: 26 Nov 2007, 21:33:03 UTC

[quote]Also, we're using the results to compare doing this work on BOINC to the BlueGene supercomputer at RPI. Who knows, with enough participants in the project -- you guys might beat out the 7th fastest supercomputer in the world! [/quoted]

If that is the truth, you will need to make some changes in order to attract more of the serious crunchers and the everyday guys that do what they can.

First the 8 WU limit needs to go. Some projects limit the number of WU's at first but as you return them your limit is increased. If I run this project exclusively on one of my Quads it goes through the 8 WU's in about 12 minutes and sits idle for another 8 minutes before downloading more WU's. This is unacceptable as it is a waste of crunching. Also it negates the practice of most experienced crunchers use of trying to keep 3 or 4 days of work available. They do this so they will allways have work even when the project runs out of work, server goes down or even if the server is being maintained on a regular basis.

Have each WU run on 2 or 3 different computers to help prevent overinflated credit from so called optimized clients.

Use larger WU's. While it is fun to watch a Quad zip through 6 minute WU's, four at a time a 1 to 2 hour WU will use less overhead and cause more efficient work.

Be generous (but not silly) with credit. A lot of projects are skimpy with credit and the crunchers with the really big systems will stay away from them because they can get more credit elsewhere. As a reference a Quad on Sudoku will get about 2200 - 2500 credits a day. This is somewhat on the low end. The same machine on SETI will get between 3000 and 3500.

Keep in mind that this is not to critisize but just some advice from a long time cruncher.
ID: 615 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Paratima

Send message
Joined: 15 Nov 07
Posts: 31
Credit: 56,404,447
RAC: 0
Message 617 - Posted: 26 Nov 2007, 22:04:22 UTC

Amen!
ID: 617 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Irishgeezah
Avatar

Send message
Joined: 10 Nov 07
Posts: 37
Credit: 11,855,733
RAC: 0
Message 621 - Posted: 27 Nov 2007, 0:30:18 UTC

The way credit is determined has to be looked at. In it's present format, older machines seem to be awarded more credit per WU than more powerful systems
I've done an approximation of 2 of my machines with the following results

1, AMD 3500+
In a 24 hour period it can crunch 216 MW WU's, being awarded 296 credit, in the same period it can crunch 6 Seti@Home WU's being awarded 320 - 450.

2, AMD 1500+
In a 24 hour period it can crunch 119 MW Wu's being awarded 199 credit in the same time it crunches 1.25 Seti@Home WU's being awarded 62 - 100.

Surely the newer, more powerful machines should earn more credit.


ID: 621 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile Travis
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist

Send message
Joined: 30 Aug 07
Posts: 2046
Credit: 26,480
RAC: 0
Message 626 - Posted: 27 Nov 2007, 3:27:15 UTC - in response to Message 621.  

Just an update on this. We've recently put in place a new validator/credit granting mechanism. We know approximately how much work any given work unit is going to be, so we calculate credit size. Right now for the current work units it'll be around 1 credit per work unit. In the future when more complex modeling code gets added credit should increase proportially to CPU time.

If you guys don't think 1 credit is sufficient for the current workunit size, i was thinking about bumping it up to 2.

As to the workunit queues, having a user buffer 2-3 days of workunits is just out of the question because of the way what we're doing works. It's not like we just have a bunch of flat data that needs to be crunched through. We're doing parameter optimization (with a lot of parameters) over a rather complex model, so we're constantly updating a population of best parameters currently found, and using these to generate new parameters for people to evaluate (using genetic search, and soon with particle swarm as well). With this in mind, the faster the turn around for work units, the more valuable they are for us - because they'll have a greater chance of improving the population and getting us to a result quicker.

That being said, a buffer of 8 workunits might not be enough. I'm trying to find a balance between fast turn around and not having results be lost on the server due to the fact we don't use a quorum. I'll try bumping it up to 20 and maybe that way machines won't be sitting around doing nothing.
ID: 626 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
zombie67 [MM]
Avatar

Send message
Joined: 29 Aug 07
Posts: 115
Credit: 501,600,404
RAC: 4,799
Message 632 - Posted: 27 Nov 2007, 6:56:28 UTC - in response to Message 615.  

Have each WU run on 2 or 3 different computers to help prevent overinflated credit from so called optimized clients.


All good stuff except for this one. That just cuts the processing power/speed by 50% or more. There are several other ways to address cheating without the hit.

ID: 632 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile Campion

Send message
Joined: 5 Oct 07
Posts: 7
Credit: 4,566,958
RAC: 1,102
Message 656 - Posted: 27 Nov 2007, 14:20:03 UTC - in response to Message 626.  



If you guys don't think 1 credit is sufficient for the current workunit size, i was thinking about bumping it up to 2.



That being said, a buffer of 8 workunits might not be enough. I'm trying to find a balance between fast turn around and not having results be lost on the server due to the fact we don't use a quorum. I'll try bumping it up to 20 and maybe that way machines won't be sitting around doing nothing.



Please consider bumping credits up to 2.

It just feels like a waste of CPU time to be granted 1.00 credits when claiming 1.46 credits. That .46 credit "loss" per unit just rubs the wrong way and while we are here for the science and not the credits, it does make it harder to be here when other favourite projects will grant more credit.

A big THANK YOU for listening and raising the quota from 8 to 20. Doing that at least will make me give Milkyway@Home a second chance as I won't have to worry about running out of work --- same situation as Bigred, quad core running out of work and being idle.


ID: 656 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile Travis
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist

Send message
Joined: 30 Aug 07
Posts: 2046
Credit: 26,480
RAC: 0
Message 673 - Posted: 27 Nov 2007, 20:31:41 UTC - in response to Message 656.  



If you guys don't think 1 credit is sufficient for the current workunit size, i was thinking about bumping it up to 2.



That being said, a buffer of 8 workunits might not be enough. I'm trying to find a balance between fast turn around and not having results be lost on the server due to the fact we don't use a quorum. I'll try bumping it up to 20 and maybe that way machines won't be sitting around doing nothing.



Please consider bumping credits up to 2.

It just feels like a waste of CPU time to be granted 1.00 credits when claiming 1.46 credits. That .46 credit "loss" per unit just rubs the wrong way and while we are here for the science and not the credits, it does make it harder to be here when other favourite projects will grant more credit.

A big THANK YOU for listening and raising the quota from 8 to 20. Doing that at least will make me give Milkyway@Home a second chance as I won't have to worry about running out of work --- same situation as Bigred, quad core running out of work and being idle.




I've bumped credit to 2 because this seems more in line with what was granted before. There are probably a bunch of WUs out there with a credit of 1 specified but as soon as these are over they should be up to 2.
ID: 673 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote

Message boards : Number crunching : So you want us to try to beat your supercomputer.

©2024 Astroinformatics Group