Processing Certian Work Units for better Cr per hour

Message boards : Number crunching : Processing Certian Work Units for better Cr per hour
Message board moderation

To post messages, you must log in.

AuthorMessage
Profile WorWizard

Send message
Joined: 13 Mar 00
Posts: 88
Credit: 8,592,514
RAC: 0
United States
Message 346786 - Posted: 23 Jun 2006, 19:47:34 UTC

Going by work unit file names. Is there any way to know which ones will give the highest credit per hour of cpu time?
ID: 346786 · Report as offensive
Odysseus
Volunteer tester
Avatar

Send message
Joined: 26 Jul 99
Posts: 1808
Credit: 6,701,347
RAC: 6
Canada
Message 346800 - Posted: 23 Jun 2006, 19:54:41 UTC - in response to Message 346786.  

Going by work unit file names. Is there any way to know which ones will give the highest credit per hour of cpu time?

No, not in general. The large differences in processing time, together with the smaller differences in earning rate, go according to the angle range—which is not part of the WU-naming schema. The detector-position info is in the header portion of each WU file, and the AR is reported with each completed result.
ID: 346800 · Report as offensive
Profile KWSN - Chicken of Angnor
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 9 Jul 99
Posts: 1199
Credit: 6,615,780
RAC: 0
Austria
Message 346805 - Posted: 23 Jun 2006, 19:57:54 UTC
Last modified: 23 Jun 2006, 19:58:07 UTC

When you open up a work unit in a text editor, you will see what Odysseus meant -

the beginning should be similar to this:
<workunit>
<workunit_header>
<name>01ap99ab.7359.32290.392322.3.123</name>
<group_info>
<tape_info>
<name>01ap99ab</name>
<start_time>2451271.0888238</start_time>
<last_block_time>2451271.0888238</last_block_time>
<last_block_done>32290</last_block_done>
<missed>0</missed>
<tape_quality>0</tape_quality>
<sb_id>0</sb_id>
</tape_info>
<name>01ap99ab.7359.32290.392322.3</name>
<data_desc>
<start_ra>23.007387279425</start_ra>
<start_dec>23.623899791967</start_dec>
<end_ra>23.036633800036</end_ra>
<end_dec>23.623899791967</end_dec>
<true_angle_range>0.41102126188338</true_angle_range>

- this is starting from line 1, it's not far down. "true_angle_range" gives you the AR of the WU in question. There really is no way to work it out by the name.

Regards,
Simon.
Donate to SETI@Home via PayPal!

Optimized SETI@Home apps + Information
ID: 346805 · Report as offensive
Profile WorWizard

Send message
Joined: 13 Mar 00
Posts: 88
Credit: 8,592,514
RAC: 0
United States
Message 346823 - Posted: 23 Jun 2006, 20:16:42 UTC - in response to Message 346800.  

Going by work unit file names. Is there any way to know which ones will give the highest credit per hour of cpu time?

No, not in general. The large differences in processing time, together with the smaller differences in earning rate, go according to the angle range—which is not part of the WU-naming schema. The detector-position info is in the header portion of each WU file, and the AR is reported with each completed result.




Ok, So, Does anyone know what would be the "optimal" angel range that produces the most credits per hour?
ID: 346823 · Report as offensive
Profile KWSN - Chicken of Angnor
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 9 Jul 99
Posts: 1199
Credit: 6,615,780
RAC: 0
Austria
Message 346831 - Posted: 23 Jun 2006, 20:29:20 UTC
Last modified: 23 Jun 2006, 20:30:22 UTC

Very much depends on your computer - no easy way to make sure.

Also, optimized clients and the standard one react differently.

For example -
on "short" WUs (going by processing time), one of my hosts using an optimized client was only 25-40% quicker than the default one.
on "long" WUs, it was actually about 200% quicker.

So the default cruncher is better at "short" (high AR or around 1.0) WUs or "long" ones than my optimized client (on my machines).

That's on one machine (Athlon64). On a P4 with SSE3, things look different again - the default is a little slower compared to the optimized ones than on my A64s, and it scales differently with ARs again.

It all boils down to "let it crunch a few WUs, divide time in seconds by credits, figure out best ratio".

If you really think it's worth the effort - you'll be spending a lot of time "fixing" your WUs, so time saved? I'm not that sure...for me, my own time is more valuable than my machine's.

Regards,
Simon.
Donate to SETI@Home via PayPal!

Optimized SETI@Home apps + Information
ID: 346831 · Report as offensive
Profile WorWizard

Send message
Joined: 13 Mar 00
Posts: 88
Credit: 8,592,514
RAC: 0
United States
Message 346936 - Posted: 23 Jun 2006, 21:53:36 UTC - in response to Message 346831.  

Very much depends on your computer - no easy way to make sure.

Also, optimized clients and the standard one react differently.

For example -
on "short" WUs (going by processing time), one of my hosts using an optimized client was only 25-40% quicker than the default one.
on "long" WUs, it was actually about 200% quicker.

So the default cruncher is better at "short" (high AR or around 1.0) WUs or "long" ones than my optimized client (on my machines).

That's on one machine (Athlon64). On a P4 with SSE3, things look different again - the default is a little slower compared to the optimized ones than on my A64s, and it scales differently with ARs again.

It all boils down to "let it crunch a few WUs, divide time in seconds by credits, figure out best ratio".

If you really think it's worth the effort - you'll be spending a lot of time "fixing" your WUs, so time saved? I'm not that sure...for me, my own time is more valuable than my machine's.

Regards,
Simon.



yea, it's very strange. and I can't put my finger on as to why. I was taking a look at the results and started to make a spreadsheeet of file name time in K and credit granted and there were some WILD differences. I wasn't really going to preform this procedure on all my cpu's. There's not enough time in the day for that. but I did want to do it on one machine just to see if it's really the case. I just divided the credit by the time and the differences were wildly spread apart on only like 20 or so work units

if your interested I can show you what I mean

ID: 346936 · Report as offensive
Profile KWSN - Chicken of Angnor
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 9 Jul 99
Posts: 1199
Credit: 6,615,780
RAC: 0
Austria
Message 346954 - Posted: 23 Jun 2006, 22:06:05 UTC

Nah,

you don't need to - I know exactly how much variance there can be.

That's why I said there's no easy answer :o) It all depends, like so much in life.

Regards,
Simon.
Donate to SETI@Home via PayPal!

Optimized SETI@Home apps + Information
ID: 346954 · Report as offensive
Odysseus
Volunteer tester
Avatar

Send message
Joined: 26 Jul 99
Posts: 1808
Credit: 6,701,347
RAC: 6
Canada
Message 346960 - Posted: 23 Jun 2006, 22:10:45 UTC - in response to Message 346936.  

[…] I was taking a look at the results and started to make a spreadsheeet of file name time in K and credit granted and there were some WILD differences. […]

If you’re running BOINC v5.2.6 or higher (as recommended), you’ll get a more accurate analysis from your claimed credit than from what you were actually granted; the latter gets thrown off on occasion, when there are hosts running older BOINC clients in the quorum. The ‘old-fashioned’ time-&-benchmarks-based claims are sometimes higher than they’re supposed to be (according to the Flop count), and sometimes lower—again, depending on the AR in each case. So the stats for claims should be at least a little less “wild” than those for grants.
ID: 346960 · Report as offensive
Profile WorWizard

Send message
Joined: 13 Mar 00
Posts: 88
Credit: 8,592,514
RAC: 0
United States
Message 346962 - Posted: 23 Jun 2006, 22:11:46 UTC - in response to Message 346954.  

Nah,

you don't need to - I know exactly how much variance there can be.

That's why I said there's no easy answer :o) It all depends, like so much in life.

Regards,
Simon.



otay. I was just doing it, just to do it. curiosity I suppose. just didn't make any sence to the fact that if I spent 8k secs oh a unit that I got the same or close to the same for spending 38k secs on a unit. odd

ID: 346962 · Report as offensive
1mp0£173
Volunteer tester

Send message
Joined: 3 Apr 99
Posts: 8423
Credit: 356,897
RAC: 0
United States
Message 347002 - Posted: 23 Jun 2006, 23:15:58 UTC - in response to Message 346962.  

Nah,

you don't need to - I know exactly how much variance there can be.

That's why I said there's no easy answer :o) It all depends, like so much in life.

Regards,
Simon.



otay. I was just doing it, just to do it. curiosity I suppose. just didn't make any sence to the fact that if I spent 8k secs oh a unit that I got the same or close to the same for spending 38k secs on a unit. odd

There are at least two reasons.

If the units errored out with a -9 error, those are definitely getting less credit than they deserve.

If the work unit didn't error out, I don't think the credits are that far out of line, but they will vary (along with the crunching time) by angle-range.

If you kill a work unit because you don't like the ratio of credit to time, then you are pushing those "low credit" work units onto others.
ID: 347002 · Report as offensive
Profile WorWizard

Send message
Joined: 13 Mar 00
Posts: 88
Credit: 8,592,514
RAC: 0
United States
Message 347171 - Posted: 24 Jun 2006, 2:03:39 UTC - in response to Message 347002.  

Nah,

you don't need to - I know exactly how much variance there can be.

That's why I said there's no easy answer :o) It all depends, like so much in life.

Regards,
Simon.



otay. I was just doing it, just to do it. curiosity I suppose. just didn't make any sence to the fact that if I spent 8k secs oh a unit that I got the same or close to the same for spending 38k secs on a unit. odd

There are at least two reasons.

If the units errored out with a -9 error, those are definitely getting less credit than they deserve.

If the work unit didn't error out, I don't think the credits are that far out of line, but they will vary (along with the crunching time) by angle-range.

If you kill a work unit because you don't like the ratio of credit to time, then you are pushing those "low credit" work units onto others.


-9 Error??

well, amongst my cpu's that are processing I noticed a varience from (approx) 5.1 credits per 1k cpu seconds all the way down .7 credits per 1k cpu seconds. Granted I only did 43 samples from various cpu's. but just for shit's and giggles, I wondered what would happen if I processed only those work units that gave a ratio of 5:1 on this cpu

if I kill a workunit it could also go to my other cpu's so I don't consider that an issue. Isn't that why there is an abort button? It's not a "push" procedure, just a "deny" procedure right? I try never to PUSH things to others if they don't want it

Also, Also among the cpus I have going I noticed that it "APPEARS" that the faster the cpu the "LONGER" the processing time to completion. Is this a fact or a conisendince <sp?>
ID: 347171 · Report as offensive
Profile Jim-R.
Volunteer tester
Avatar

Send message
Joined: 7 Feb 06
Posts: 1494
Credit: 194,148
RAC: 0
United States
Message 347564 - Posted: 24 Jun 2006, 14:37:19 UTC - in response to Message 347171.  
Last modified: 24 Jun 2006, 14:38:36 UTC



-9 Error??

This is the error code returned for a "noisy" work unit. There is nothing actually "erroring out" but it's just the work unit has too many "signals" in it. This can be due to interference at the receiver such as a thunderstorm, etc., a stray transmitter of some type near it, or other reasons. It is known that a work unit with a very large number of "signals" or "hits" is due to this interference so a limit is put on the number of actual signals detected before a unit is rejected due to the 'noise". (I think it's set at 30.)

if I kill a workunit it could also go to my other cpu's so I don't consider that an issue. Isn't that why there is an abort button? It's not a "push" procedure, just a "deny" procedure right? I try never to PUSH things to others if they don't want it
No it couldn't. Unless you have separate accounts for all of your computers! There is code in the scheduler that checks for this, so yes, if you abort an "undesired" wu then you *are* "pushing" it onto someone else.

Also, Also among the cpus I have going I noticed that it "APPEARS" that the faster the cpu the "LONGER" the processing time to completion. Is this a fact or a conisendince <sp?>

While I haven't noticed this myself (I have only one slow computer) it has been mentioned in many other threads. It is probably due to the specific computer architecture you are looking at. Variations in memory bandwiths, cache sizes, and many other factors come into the picture when you compare different computers, not just the clock speed. Theoretically, a 1ghz comp should crunch exactly twice as fast as my 500mhz, and a 3ghz processor should be exactly six times as fast, but if you look at actual results, this is not the case. This would not even be the case looking at two identical 500mhz computers just like mine since my comp is using 100mhz memory speeds while others may use 133mhz memory. So the same identical motherboard with the exact same *amount* of memory can have varying crunching speeds due to simply the *type* of memory installed!

While this is strictly my own opinion, I feel that anyone who "chooses" particular work units is like the person at a party that picks all of the *good nuts* out of a bowl of mixed nuts. Sure, they will all get eaten eventually, but for every "good nut" that person specifically picks, that's one less that someone else will get! As I say, this is my opinion, but everybody should have the opportunity to get an equal "mix". Put yourself in the position of the "other person" for a minute. If I were choosing all quick running or high paying work units while your computer is left slugging away for hours for low paying wu's, how would that make you feel toward me?

My opinion in this has been stated in other threads. Yes there is a "penalty" for this in place now, but it is in my opinion not strict enough. Currently you loose one work unit in allowable quota for every canceled (or errored out) work unit. While I agree with this for errors, as many errors are not due to the user themselves, I think that the penalty for purposefully aborting wu's should be bumped up to at least a reduction of 10 for every "user aborted" wu. This would discourage people eating all of the "good nuts" and leaving the "bad nuts" for others.
Jim

Some people plan their life out and look back at the wealth they've had.
Others live life day by day and look back at the wealth of experiences and enjoyment they've had.
ID: 347564 · Report as offensive
Profile MJKelleher
Volunteer tester
Avatar

Send message
Joined: 1 Jul 99
Posts: 2048
Credit: 1,575,401
RAC: 0
United States
Message 347736 - Posted: 24 Jun 2006, 17:44:27 UTC - in response to Message 347564.  

My opinion in this has been stated in other threads. Yes there is a "penalty" for this in place now, but it is in my opinion not strict enough. Currently you loose one work unit in allowable quota for every canceled (or errored out) work unit. While I agree with this for errors, as many errors are not due to the user themselves, I think that the penalty for purposefully aborting wu's should be bumped up to at least a reduction of 10 for every "user aborted" wu. This would discourage people eating all of the "good nuts" and leaving the "bad nuts" for others.
While in theory I agree with the principle, I don't think it would have any great effect. Abort one result and get the quota knocked down by 10, then quickly return a fast result and the quota gets doubled up to the max. The quota penalty just isn't designed as a discouragement for aborting work, it's to rein in the machines that have gone sour and aren't returning good work at all. Someone would have to build a separate track of work aborted and returned to affect the quota, distinct from the machine errors, for an effective deterrent. Anybody that riled up about it to tackle the job?

MJ

ID: 347736 · Report as offensive
EricVonDaniken

Send message
Joined: 17 Apr 04
Posts: 177
Credit: 67,881
RAC: 0
United States
Message 347754 - Posted: 24 Jun 2006, 18:17:17 UTC

Clearly, the only effective punishment for "cherry picking" is exponential to the same degree as the reward for correctly finishing a task.
So if you voluntarily abort a perfectly good task just because for some reason you don't like it, your quota should be halved.

OTOH, to make this issue even more complicated there are valid reasons for voluntarily aborting an WU.

I've killed a few that were apparently stuck in a infinite loop, for instance.
Worse, those units later completed just fine on someone else'e host.

Unless we have some at least close to foolproof method to detect "cherry picking", it would be =worse= to implement any punishment than to leave things as they are.

ID: 347754 · Report as offensive
Alinator
Volunteer tester

Send message
Joined: 19 Apr 05
Posts: 4178
Credit: 4,647,982
RAC: 0
United States
Message 347758 - Posted: 24 Jun 2006, 18:33:56 UTC - in response to Message 347754.  
Last modified: 24 Jun 2006, 18:35:41 UTC

Clearly, the only effective punishment for "cherry picking" is exponential to the same degree as the reward for correctly finishing a task.
So if you voluntarily abort a perfectly good task just because for some reason you don't like it, your quota should be halved.

OTOH, to make this issue even more complicated there are valid reasons for voluntarily aborting an WU.

I've killed a few that were apparently stuck in a infinite loop, for instance.
Worse, those units later completed just fine on someone else'e host.

Unless we have some at least close to foolproof method to detect "cherry picking", it would be =worse= to implement any punishment than to leave things as they are.


I'd have to agree with you here, Eric.

I have been scanning a large number of hosts since the SE rollout, and it appears for many folks who have always been stock app crunchers their output rate overall is virtually unchanged from the standard app.

Anecdotal evidence is that when a forum "newbie" posts it's more like, "Gee, why is it taking 12 hours max instead of 4 for these results?", rather than "OMG, my RAC tanked, SAH sux!".

Obviously, if you were an optimizer your RAC took a pounding, even with running Crunch3r's 5.12 (if you were lucky even to get it when it was available) or not.

Is there still some discrepancy in value between the various AR's? Yes.

Will it get straightened out eventually? In all likelyhood, yes.

Does this mean there is any advantage to "waste" time cherry picking (or preventing it)? Well, probably not, but that depends on how fussy someone is about RAC, Total Credit, Instant Gratification and/or perceived cheating. :-)

Alinator
ID: 347758 · Report as offensive
Profile WorWizard

Send message
Joined: 13 Mar 00
Posts: 88
Credit: 8,592,514
RAC: 0
United States
Message 348085 - Posted: 25 Jun 2006, 4:22:59 UTC

Like I said before. Just curious. I didn't know about the abort quotas thing. but still only curious how things workes and why
ID: 348085 · Report as offensive

Message boards : Number crunching : Processing Certian Work Units for better Cr per hour


 
©2024 University of California
 
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.