Some background and information on how HPF2 works on BOINC.

1) The goal of HPF2 is to generate about 33,000 structures predictions per gene. These predictions are then further analyzed by the researchers to identify the mostly likely structures for the protein. See the official description of the project for more info: http://homepages.nyu.edu/~rb133/Human_Proteome_Folding_Project.html

2) Each result returned by a computer contains between 10-35 predictions depending on a rough estimate of how tough the predictions will be to compute for the gene.

So on a given gene if we ask each computer to compute 25 structures, then it will take 1,320 results in order to generate the set of 33,000 structures. We could create one workunit for the gene and send it out enough times to make sure we get 1,320 results. However, with the way that BOINC works and this would create some inefficiencies in the system.

What we do is that we create a set of workunits that will generate the required number of results for the gene. Over the entire set of workunits for the gene, we need to generate an average of about 17.2 valid results per workunit. In this example where 25 results are being generated per result, this means that we will create 77 workunits for the gene - each of which needs to average 17.2 valid results with 25 structures per result. 77*17.2*25 = 33,110 structure predictions.

If we have a gene that is tougher and only 10 structures will be generated per result then we would create 192 workunits. 192*17.2*10 = 33,024 structure predictions.

In order to generate the average of 17.2 valid results per structure we only have three variables to modify:

1) Initial replicas sent out (initial_replicas)
2) The min quorum (min_quorum)
3) The min results in agreement within a quorum before accepting the result as valid (min_agreement)

The way that BOINC works is that for each workunit, BOINC will initially send out 'initial_replicas' initial copies of the workunit to be processed. As the results come back in, BOINC will wait until 'min_quorum' is reached before attempting to validate the workunit. During this phase, if a result is aborted or returned as an error, then an additional copy will be sent out. Once 'min_quorum' results are returned then validation is attempted. Validation is successful if at least 'min_agreement' results are determined to be valid. If there are not at least 'min_agreement' valid results, then all the results are marked as 'INCONCLUSIVE' and an additional result is sent out. Each time an additional result is returned, validation is attempted again. Once we have min_agreement valid results then the results will be marked as valid or invalid as appropriate and credit is awarded. Normally at this point BOINC would go ahead and 'assimilate' the result (this means to copy the results off to be returned to the researchers). However, we have modified the assimilator for HPF2 so that it will wait until the last result is returned or the result with the latest deadline has missed its deadline. Then the assimilator will collect all valid results returned to return to the scientists. Also - once the validation has been successful, no additional results will be sent out even if additional errors or invalid results are returned.

We currently have initial_replicas set to 19, min_quorum set to 15 and min_agreement at 13. This is yielding about 17.6 valid results per workunit. We want to keep it a somewhat above the 17.2 target to ensure that we have the required set of structure predictions.

Each result returned by the project contains unique data. We have some additional information in the results that we will use to help us make sure that the very strong majority of results returned contain the correctly computed data.

We are starting out sending 19 results per workunit and validating when 10 are returned **. Each workunit will run for about 8 hours on each computer. A computer running this project will generate an additional structure prediction in about 90 seconds of execution time on a average computer.

This means that if you have a really powerful computer, you may generate 1000 structure predictions in 8 hours while a slower computer may only generate 160. We will be awarding credit by averaging the the claimed credit per structure prediction, and then taking that value and awarding the # of structure predictions in a result * the average claimed credit per structure prediction. So in this case if the slower computer had a claimed credit of 80 and the larger computer had a claimed credit of 450, then the average credit per structure prediction would be 0.475. This means that the slower computer would be granted 76 credits and the faster computer would be awarded 475 credits.

The key point though is that all computers should have a estimated time to completion of around 8 hours.