For information (and only related to the current receptor):

OPNG has been tackling the current target since the end of November 2021. If the tasks I have been receiving form a fair sample of what everyone else was getting, there were no tasks with huge job counts until the 11th of January, and from that point on the only low job counts were for retries!

The new, bigger, tasks have lots of ligands with either fewer atoms, lower branch counts or both. That causes the workunit builder to think the individual jobs will run a lot quicker (I wish...) so we get a lot more of them!!!

As for 10 times as many jobs - true if one takes the extreme cases for both low and high counts! From my relatively small sample of about 2600 tasks for this receptor, I've seen less than 100 tasks with a dozen jobs or less and less than 40 with 100 jobs or more.

If it's of any interest I'll post some basic statistics related to the above (though that's likely to get very long...); I can also go back through previous target receptor data to find other cases of the many versus the few... Whatever, I hope the above helps...

Cheers - Al.

The jogs inside the WUs have beem increased substantially. Go to your Results page here and click on a completed recent and also a couple weeks older WU and you'll see the Jobs # difference in the file.

When I posted up-thread I forgot to include a link to the post by Keith (Uplinger) discussing the work unit packaging algorithm.

Unfortunately, it appears that ligands of a similar size and number of branches tend to be clustered in the data provided for work unit generation. So instead of work units getting a mix of jobs across the range of possible effort required, there are either a few large/complex ligand jobs or lots of small/simple ligand jobs... Then, unfortunately, we get the swings in run time and, in some cases, tasks failing because of limits being exceeded :-( -- during the great OPNG job storm last year, we saw all sorts of tasks at once so it tended to hide the swing effect.

I have some monitoring software which collects OPN1/OPNG job parameters per task and can confirm that (at least for my two systems[1]) it appears that tasks are [still] built using that algorithm, or something very similar. It also appears that the majority of tasks for most receptors have had job counts in the lower ranges. That would be expected if jobs of similar potential difficulty tend to get clustered... To date, I've got data for 27108 OPNG tasks (over a variety of target receptors), comprising 986498 jobs. That's an average of just over 36 jobs/task, but the actual distribution is heavily biased towards tasks with a lower job count. However, using the WCG job assessment algorithm to score the individual jobs, over 55% of them scored so low that they would end up in 75+ job tasks if grouped only with their peers!

If the scientists were able to tweak the ordering of the ligand selections per batch, this might improve things... However, that may not be as straightforward as it sounds! And there are other issues with the algorithm and [as a side-effect] credit assignment anyway; hopefully, this can all be revisited after the transition to Krembil is completed (see Igor Jurisica's post in response to a thread about low credit...)

Cheers - Al.

[1] I only have a single 1050Ti and a single 1660 Ti in a pair of Linux machines and don't run anything to try to maintain a continuous stream of OPNG tasks, so the statistics I see won't have as many of the really extreme cases as might show up for a big hitter!