There's a facility in BOINC called homogeneous redundancy that allows work management based on various aspects of host systems, including the Operating System (as observed here) and/or specific aspects of the actual processor in use.

The theory is that if validation is only against similar hardware on similar operating systems there should be less validation errors resulting from [subtle] differences in calculation outcomes, because of differences in hardware, libraries or generated code.

In some cases it's essential - ARP1 is an example, because bit-wise identity of returned results is required, and the best chance of that is to have both a single O/S and "bit-ness" for all tasks for a given work unit (32-bit and 64-bit binaries use different floating point instruction sets...) In other cases it might be less necessary, depending on how validation is done...

As far as I know, WCG always does this for CPU-based projects.

Hope that helps.

Cheers - Al.

Adri,

I obviously didn't go into as much detail as I could have done about how homogeneous redundancy works :-) [1]

I sometimes get MCM1 retries for work units where the first two both went out to Windows systems but they returned results with a non-success state close enough together that the generated retries became available to any platform because, as you noted, the forced platform selection does not apply if no tasks are still out in the field (and adaptive replication [as per your example] is probably the most common reason for platform switches...)

Under normal circumstances I tend to see these when the Windows tasks go to Windows systems that fail with some variant of "can't start process", but it became somewhat more common when the download problems were at their height :-)

This sort of stuff really belongs in a set of new [or reorganized] FAQs -- Sekerob and others did good work 12 or more years ago, but much has changed since then and some weeding out and re-writing of that stuff might be a future project(?) if only we can be sure that people would even look there in the first place! :-)

Cheers - Al.

P.S. In my earlier post I could/should also have noted that it is the reason we often see messages that tasks are committed to other platforms when doing work requests :-)

[1] Quoting large sections of the BOINC Wiki entry didn't really seem like a good use of space for a single post :-)

What type of errors? Could you link to some of the error tasks?

Have you checked for hardware errors, e.g. run me test etc?