Message boards : Number crunching : Crunch results sideways
Message board moderation

To post messages, you must log in.

AuthorMessage
Profile Andy Gray
Avatar

Send message
Joined: 28 Sep 04
Posts: 9
Credit: 801,351
RAC: 39
Message 9852 - Posted: 2 Sep 2005, 15:12:17 UTC

Hey Ziran, simple but smart. Can't fault the logic, but can their system cope?
ID: 9852 · Report as offensive     Reply Quote
Travis DJ

Send message
Joined: 29 Sep 04
Posts: 196
Credit: 207,040
RAC: 0
Message 9859 - Posted: 2 Sep 2005, 17:24:45 UTC

I prefer the *chaos* method. It's pretty simple. The server will create all the workunits and then distribute them in a totally random fashion. This will ensure an entirely uneven distribution of work! This way work will complete with the longest time investment possible and may even force the admins to increase the workunit deadline time!!

As the Guiness man says: Brilliant!

ID: 9859 · Report as offensive     Reply Quote
Profile Lee Carre

Send message
Joined: 13 Jul 05
Posts: 23
Credit: 22,567
RAC: 0
Message 10489 - Posted: 28 Sep 2005, 12:11:03 UTC - in response to Message 9834.  

<blockquote>Most projects want the individual results back as soon as possible, but LHC is different, it wants the whole batch back as soon as possible. That makes the individual turnaround time for each result and WU unimportant. What matters is the return time for the last wu to form a quorum of the last result. So then other projects try to crunch one result at a time, Wouldn't it better for LHC to crunch the results sideways? Let me explain what i mean by this.

The difference is in what order the work is sent out.

If you send out one result at a time, like we do today, you send out all 5 WU's for the first result, then 5 for the second result, the third...

It would look like this: 1,2,3,4,5,1,2,3,4,5,1...

If you crunch results side ways you send out one WU from all results before you send out the second WU from any result.

It would look like this: 1,1,1...2,2,2...3,3,3...4,4,4...5,5,5...

The strength of sending one result at the time is, that the time from the sending of the first WU from the result, till a quorum is formed, is minimised. This is important for some projects. The drawback is that it is a micro optimisation, focusing on one result at the time.

The strength of doing it sideways is that WU's from more results can be crunched on simultaneously. This will also increase the time from the sending of the first WU from a result, till a quorum is formed, but decrease the time from generation of work, till the sending of the first WU.

How will this speed things up? Since work on the batch is spread more evenly over time, it might not be needed to send out all WU's for all results. Some results will formed a quorum before the 4th WU is sent out, making it unnecessary to send out NR 4 and 5. Even more will form a quorum before WU 5 is sent out, making NR 5 unnecessary.

Since the first WU of all results are sent out faster, they will also have an earlier deadline. if a deadline is missed, a replacement can then be sent out sooner.
</blockquote>

seems like a sensible idea to me ;)

Lee
ID: 10489 · Report as offensive     Reply Quote
Gaspode the UnDressed

Send message
Joined: 1 Sep 04
Posts: 506
Credit: 118,619
RAC: 0
Message 10491 - Posted: 28 Sep 2005, 13:41:59 UTC

Judging by the way results have been issued, and the number of questions about why only one result has been issued for a WU, it seems they're doing this already.

Nice idea, Ziran, but I think LHC@Home beat you to it.


Gaspode the UnDressed
http://www.littlevale.co.uk
ID: 10491 · Report as offensive     Reply Quote
itenginerd
Avatar

Send message
Joined: 29 Aug 05
Posts: 42
Credit: 27,102
RAC: 0
Message 10494 - Posted: 29 Sep 2005, 3:16:41 UTC - in response to Message 10491.  

<blockquote>Judging by the way results have been issued, and the number of questions about why only one result has been issued for a WU, it seems they're doing this already.

Nice idea, Ziran, but I think LHC@Home beat you to it.

</blockquote>

Now if only they could get to the end of the friggin' list and send the next results for the WU's I've got a stake in.... 8)

(j)
ID: 10494 · Report as offensive     Reply Quote

Message boards : Number crunching : Crunch results sideways


©2022 CERN