ARM BRP4 app version invalids
24 Jun 2022 12:26:44 UTC
Topic 227715
(moderation:
I've seen an uptick lately of invalid BRP4 results, with all of them being when there are different versions of the BRP4 ARM app. It appears that there is a higher rate of inconsistent results when a 1.46/1.47 task and a 1.60/1.61 task are in the same WU, and then it is luck of the draw of which version gets the third task for confirmation. If it is a 1.4x version, those two will be valid and the 1.6x invalid, and vice versa.
Not sure that there is anything that can be done (and, full disclosure, I'll probably be jumping my Androids back to WCG when they are back online, so this feels a bit like criticize and run), but with the refocus on BRP4/7 analysis I figured I would bring it up in case it is a pervasive issue.
It is a common issue. You
)
It is a common issue. You see it on consensus between AMD, Intel and Nvidia cards for other apps and projects also.
Thanks for confirming that it
)
Thanks for confirming that it isn't a BRP4/ARM issue. Quite unfortunate that it is common and persists; seems like a waste of precious computing power.
The BOINC goal is to have any
)
The BOINC goal is to have any project stay below 5% of invalids-errors. All the projects achieve that as AFAIK.
You will never achieve 0%-0% likely on any project. Most are in the 2.5 - 3.5% range.
Maybe I'm an outlier on this
)
Maybe I'm an outlier on this or it is simply due to my very small sample size, but BRP4 is currently a bit over that ideal rate.
When looking at invalid, inconclusive, and valid, I'm seeing 13.5% invalid (7/52), 15.4% inconclusive (8/52), and 71.2% valid (37/52). All of them are due to the app version mismatch. Unfortunately, I haven't been keeping too close of an eye on things as I've just been parking a couple of Androids here while WCG is down. I think that this is a recent increase in invalid rates, but can't say for sure.
You should not include
)
You should not include inconclusive. Those are tasks that have not yet been validated as either valid, invalid or errors. Those inconclusive are simply waiting for a wingman result or have yet been pushed through the project validation server.
You have hidden your computers so I can't look and calculate your rate for you.
The project made a change to which computers get BRP4 tasks lately. Most hosts have been moved to BRP4G tasks. Only the slow Androids or ARM SoC's mainly are still crunching BRP4.
Owing to the math precision differences among platforms is the main reason why you see invalids among version mismatches.
I'm guessing it is the change
)
I'm guessing it is the change of BRP4 allocation that you mention that has resulted in the increase in invalids from app version clashes. As the ARM output is basically a rounding error in the compute power of the project, and no one else has jumped on this thread, I think it is time to let it rest. Thanks for your engagement with the topic.
On a side note and to clear up some confusion, the figures that I included above as inconclusive are not those that are pending validation (quorum 2 WU waiting on the second result, "Completed, waiting for validation"), they are those that have already reached quorum 2 and those two results don't match ("Completed, validation inconclusive"). So they have already experienced the app version mismatch and are awaiting a third result, which will cause one of the first two to be marked as invalid. The total I used above (52) was the total of those three states and didn't include tasks in progress or pending validation; so the figures are saying "Of the WUs that have reached quorum 2, only 71% are valid."
Of course, these numbers improve if we were to look at all of the tasks within the WUs, and not just from my perspective (111 total tasks in the WUs above, 7 invalid, 16 inconclusive, and 88 valid). And this is probably what we should be concerned with--it pushes the invalid rate down to 6.3%, which is far more acceptable.
Here are the full numbers from my account. They've changed slightly since I first posted, but the ratios are consistent:
Total: 73
In Progress: 8
Pending validation: 14
Validation inconclusive: 8
Valid: 36
Invalid: 7