Prime95 v30.3

[chalsall]

Originally Posted by S485122

But why should "ordinary" users of PrimeNet be forced, by default, to participate in that work ?

A valid observation and argument.

Please keep in mind this is now in beta testing. This is where all the early adopters help "shake-out" the code, and find "edge-cases".

George has already said that there won't be enough CERT work to give to those who only want that kind of work.

This kind of work would actually be of interest to people like me, who run GIMPS on the kit we're responsible for to do ongoing sanity checking. CERT runs are shorter than DCs, so would point out suspect kit more quickly.

[Batalov]

Would it make sense to fund and build a specialized on-premises server (or a small AWS farm) which will _only_ run "CERTs"?
That would help to avoid very large data traffic back and forth. (Just process it where data resides.)

Could some back of the envelope estimates be run? How many cores would simply take care of the time averaged queue of CERT verifications, nearly real-time.

[chalsall]

Originally Posted by Batalov

Could some back of the envelope estimates be run?

You're entering what's known as "Analysis Paralysis".

We have code. We have compute.

Run with it, and see what happens.

[ATH]

There are many options: When people choose their primary work type they could be presented with options for secondary work type PRP CERT and/or PRPCF CERT or NONE.

It could also be a primary work type: PRP CERT or PRP+PRPCF CERT and then as a secondary option you choose what work type to run when no CERTs are available.



I agree that the "normal" / "average" GIMPS user should not be forced to do PRP CF work even just short CERT work.

[Xyzzy]

Originally Posted by Prime95

What do others think? I could easily change the default to only cert exponents over 50M unless you've signed up for PRP-CF work. It would be more work, but I could also try handing out PRP-CF certs in batches of 5 or 10 for less interruptions.

Originally Posted by Xyzzy

If there was a cert worktype option we are certain enough people would opt-in to take care of them.

[Viliam Furik]

I am not sure if somebody didn't already mention it, but if there was a checkbox (disabled by default) whether the user minds to also run CERTs on the machine, I think that there would be enough people to do CERTs, and it would not be an only-cert work type, so no issue with a no-work-to-do problem.

[Batalov]

Originally Posted by chalsall

We have code. We have compute.

Run with it, and see what happens.

Sheesh! That's what already happened - it had been run and you got 20+ complaint messages. :-)

Time to think? Or to continue sawing off the branch that you sit on?

Also, a month has not finished yet, and people didn't receive their internet summaries/bills yet.
One or two people might hit their limits after which internet bandwidth is throttled by the provider. If one downloads 100Mb every 30 minutes, how soon will they hit this limit? What if they also have a habit of watching some netflix or whatnot?

I have once downloaded a Tb worth of public genomes so I know - this limit does exist (and it is indeed spelled out in the AT&T's Uverse contract), test it yourself, if you want. (and I am talking about an expensive plan, other people's limits might be less that 1Tb / month

[chalsall]

Originally Posted by Batalov

Sheesh! That's what already happened - it had been run and you got 20+ complaint messages. :-)

George is getting empirical data, and listening to the various perspectives of those running his ***beta*** code.

Doing a cost/benefit analysis of running a dedicated CERT farm makes absolutely no sense (IMO).

[pinhodecarlos]

I don’t mind doing only CERT work, if you can force my machine to do it on the back please do it so.

[jwnutter]

Originally Posted by chalsall

Doing a cost/benefit analysis of running a dedicated CERT farm makes absolutely no sense (IMO).

Assuming I'm following this post correctly, wouldn't it make more sense to use a dedicate server farm to complete the remaining DC work? This would at least allow someone to calculate a potential end-point for such a cost/benefit analysis. And, it would give the project time to determine new workflow patterns once this version is released to the public.

[preda]

Originally Posted by Batalov

Could some back of the envelope estimates be run? How many cores would simply take care of the time averaged queue of CERT verifications, nearly real-time.

First, about the size of CERT: a cert assignment is *one residue* in size (download), thus much smaller than a Proof file (which is 10x residue in size, upload). At the wavefront one residue is 12MB.

Assuming 1000 PRPs completed daily, and the CERT requiring 1/200 the work of the PRP, we come with needing capacity equivalent for running 5 PRPs per day.