Workunits with a quorum of one?
11 Jan 2018 15:11:12 UTC
Topic 212602
(moderation:
When I noticed that my little I-3 powered Linux box was screaming through tasks in less than 3 hours instead of the usual 7, I looked at my results on the Web site to see what was wrong. I was surprised, but pleased, to see that they were all completing normally.
What I found more interesting was that most of them were run with a quorum of just one, with no partner.
I had assumed that tasks would always be run by at least two participants as a check against computational errors.
What is it about these tasks that they're run without this check?
It's a BOINC feature called
)
It's a BOINC feature called 'adaptive replication' or something like that. Basically it means that if a host has a good enough record to be 'trusted', the scheduler will accept the result without sending a duplicate to a second host for confirmation. It first came in at this project for FGRP CPU tasks about 18 months ago. You can read an interesting discussion about it here.
This is not as 'risky' as you might think at first glance. A fairly strict 'sanity' check is made of all results before being accepted or before being submitted to a direct comparison. A single 'bogus' result could easily fail the sanity check if whoever/whatever created/caused the bogus result wasn't aware of what is needed for a result to be declared 'sane'. I haven't heard of any problems arising from trusted hosts being allowed to submit results that auto-validate after just the sanity check.
For 'technical' questions like this, you're more likely to get more discussion (and more quickly probably) if you post in a 'technical' forum like Cruncher's Corner rather than the Cafe :-).
Cheers,
Gary.