Questions and Answers : Wish list : Categories of crunchers
Message board moderation

To post messages, you must log in.

AuthorMessage
arturg

Send message
Joined: 29 Sep 05
Posts: 10
Credit: 11,373
RAC: 0
Message 12501 - Posted: 27 Jan 2006, 14:08:52 UTC

Yesterday Keck_Komputers told me that 4th and 5th result are not needed, after quorum is completed what was very strange to me, because it means that sometimes we could lose about 40% of work and I started to think, that maybe good idea would be creating categories of crunchers and combine it with method of sending results. It's just idea, mayby wrong (??).
Proposed categories:
-6 hours cruncher (or 8, 10?),
-1 day cruncher
-1 week cruncher
-"not descibed"
If sombody declares himself as 6 hours cruncher, he make efforts, to crunch every result during 6 hours. It is not only ipmortant, that he crunch each result for max.6 hours, but that he declares himself as such user, what additionally makes his quickness more sure.
Maybe some categories deadlines would be nessesery.

If we have categories, we could send results with better certainity, that they will come back during category time:
-3 results of WU to 3 users of category 6 hours
-3 results to 3 users of category 1 day
-4 results to 4 crunchers of category 1 week
-5 results to 5 crunchers of category "not descibed" (probably it is the method used today)

In 6 hours category - 4th of 5th result could be distribute only, when quorum is not completed during 6 hours.
In one week category 5th result could be sent to 6 hours user, after quorum is not complited after 1 week.

Additionally each user should know, that he could crunch on various way, which is best for his possibilities and that he works with better efficiency for science if he choose category.

Artur


ID: 12501 · Report as offensive     Reply Quote
arturg

Send message
Joined: 29 Sep 05
Posts: 10
Credit: 11,373
RAC: 0
Message 12504 - Posted: 27 Jan 2006, 14:32:56 UTC

Additionally advantage of such categories would be, that everbody (in most of cases) would have so quickly credits as quickly he send results.

Artur
ID: 12504 · Report as offensive     Reply Quote
Profile Keck_Komputers

Send message
Joined: 1 Sep 04
Posts: 275
Credit: 2,652,452
RAC: 0
Message 12517 - Posted: 27 Jan 2006, 21:07:00 UTC

Even though the extra redundancy is not needed for the science it is needed for the turnaround. Turnaround is very important at this project since the next group of studies is often based on the previous group.
BOINC WIKI

BOINCing since 2002/12/8
ID: 12517 · Report as offensive     Reply Quote
arturg

Send message
Joined: 29 Sep 05
Posts: 10
Credit: 11,373
RAC: 0
Message 12529 - Posted: 28 Jan 2006, 0:53:20 UTC - in response to Message 12517.  

Even though the extra redundancy is not needed for the science it is needed for the turnaround. Turnaround is very important at this project since the next group of studies is often based on the previous group.


That's right only if we look at one WU but it's enough to take 2nd WU and this turnaround for both WU could be better, when we better manage distribution. Even when we don't have categories and use quorum 3 of 5 as today we could receive better turnover, when we distribute not random. Very simple example: we have 3 quickly users an 7 slowly 1 week users. In that case only when we place quickly users in the same WU team we have better turnover (1 unit quickly and 1 slowly). In all other cases we must wait 1 week for both results. Any failers are in this case not important in relation to today's distribution. When we have more users it's very clear that combinig slowly users with quickly makes whole turnover slower. Each case QQWWW or QWWWW is in 100% lost of turnover. The significance depends of course of whole number of Q users and W users, but if we can see our peding credit tables - number of W (1 week users) is high. On the other hand QQQQQ team is not good for turnover, because we lose quickly turnover-producer.
On the other hand even if 1 of 15 results has client error and we distribute QQQ QQQ QQQ QQQ QQQ and not QQQQQ QQQQQ QQQQQ at the same time we have 4 WUs crunched (+5th partially) and not 3.

Artur

ID: 12529 · Report as offensive     Reply Quote
arturg

Send message
Joined: 29 Sep 05
Posts: 10
Credit: 11,373
RAC: 0
Message 12542 - Posted: 28 Jan 2006, 10:24:42 UTC

Take a look at this 2 extreme ways of distribution:

Let's take more users for example 600 quickly (Q), which need 6 hours for calculation of WU and 300 slowly(W), which need 1 week for caculation. So we have good team.
Let's take bigger portion of 900 results to calculate in 300 WU and quorum is 3 of 3.

If we distribute by category (or just by speed of reporting results):

200 work units QQQ and 100 work units WWW

we can receive 200 WU after 6 hours and 100 WU after 1 week,

but if we have malignant distributor, which makes 300 mixtures QQW we receive this whole portion of 300 WU after 1 week.

We assume, that failures are independent of category.

So majority of work (200 WU) is done after 6 hours and not after 1 week.
The difference is significant.

Artur

ID: 12542 · Report as offensive     Reply Quote

Questions and Answers : Wish list : Categories of crunchers


©2024 CERN