GIMPS future

[Complex33]

I am curious as to what GIMPSters think is the best path for the project to take. Any and all suggestions would be appreciated, whether they deal with the server, the type of work to be done, the client, etc. Any and all ideas are welcome. I envision this to be an open thread of long term goals for the GIMPS project. So speak up and let’s hear what the most important sector of this project thinks, its users.

[outlnder]

I would like to see 2 phases of the project get a little closer together.

IMHO, double checking seems far behind of LL testing. Too many bad results are coming in from over tweaked machines and the owners are not aware of it. This could be just my problem, but I feel a LL test should be double checked within 6 months of completion. This would ensure fairly quick feedback to an individual of positive results from their machines.

I think a system to notify the owners of a machine that it's results are bad should be set up. No one wants to see their hard work begin to disappear in 1 to 2 years from now.

I also would like to see clients built for each CPU type. I know George does this for fun, but he does get immortallized everytime Gimps finds a new prime. Perhaps he could get some other programmers to build a client for more machine types.

[edorajh]

Originally posted by outlnder

I think a system to notify the owners of a machine that it's results are bad should be set up. No one wants to see their hard work begin to disappear in 1 to 2 years from now.

I think that is important. Not sure how this could be done, but some kind of system that gives early notification about error producing machines will be very helpful.

[Complex33]

Originally posted by outlnder
I would like to see 2 phases of the project get a little closer together.

I agree with you and that is why I like the idea of a new user's install, starting with DC's, both to give hardware feedback and get a faster stats fix. I'd like to see that implemented in the program itself. Along the stats line, is there a way that the program could send users through the help menu to their stats, the fourm, etc?

[PrimeCruncher]

Originally posted by Complex33
Along the stats line, is there a way that the program could send users through the help menu to their stats, the fourm, etc?

I don't see why not. It should be very easy to program.

Personally, I'd like to be able to sort reserved exponents on the individual stats report by column, i.e. so you can immediately see which ones are close to expiring, how many on one computer, etc. which could be a boon to people with large farms or those who have put Prime95 on large amounts of machines (I only have 10 machines and I have 90 exponents reserved!). It would also be nice if you could change your info (i.e. Name, E-maill addy, etc.) from that page as opposed to having to do it through the client.

[aaronl]

Originally posted by outlnder

I also would like to see clients built for each CPU type. I know George does this for fun, but he does get immortallized everytime Gimps finds a new prime. Perhaps he could get some other programmers to build a client for more machine types.

The overall gain decreases pretty quickly as you add additional machine types. PowerPC is the obvious choice because it's the second most popular architecture among personal computers. However, I don't think AltiVec helps because as far as I know it's limited to single precision types. Rewriting the client for another chip is a lot of work, and I'm not expecting anyone to do that for a niche platform (even PPC) anytime soon. Opteron/AMD64 is a special case because it's backward compatible. George is already working on it.

I wish Itanium wasn't so expensive and rare because it sounds very capable. George mentioned that the Itanium system he was using to double check M(40) is outrunning his P4 using a program writen in C. Of couse, the Itanium system probably costs an order of magnitude more.

[smh]

I wish Itanium wasn't so expensive and rare because it sounds very capable. George mentioned that the Itanium system he was using to double check M(40) is outrunning his P4 using a program writen in C. Of couse, the Itanium system probably costs an order of magnitude more.

IIRC it's a quad processor Itanium

[aaronl]

How can you parallelize the LL test?

[gbvalor]

Actually, the verification is running in a quad Itanium-II machine but using only TWO processors.

Using HP compiler for Itanium, Glucas squeezes even more these chips. A two threaded version of Glucas, using two Itanium-II at 900 Mhz run at about 0.045 sec/iter for a 1024 FFT runlength. Obviously, those machine are good, but no cheap. Compare it with 0.100 sec/iter in GIMPS 1.400 Mhz opteron for a single-threaded version of Glucas.

Some time ago I run a 16 threaded version (I think they were alphas). The efficience per processor droped about 30% from a 8-processor version. After my few experiences in so kind of systems I think that 8 is a reasonable limit.

If GIMPS were hired such machines we likely would finish a verification in no more than three of four days . But perhaps we need more time to prepare the announce and the party ...

Guillermo

[GP2]

Guillermo,

Have you considered contacting Clearspeed and asking them to send you the SDK for their product?

From reading their FAQ, the SDK bundled with hardware development board costs $25K, out of the question. But the SDK alone (without hardware) has an instruction set simulator that allows benchmarking. They have a C compiler for it, it would be very interesting to see how Glucas benchmarks.

It's not a consumer-level product, but there would certainly be a distributed-computing hobbyist market for it if growing your farm meant adding a PCI card instead of a whole new box, and with negligible power consumption.

[Xyzzy]

Originally posted by GP2
It's not a consumer-level product, but there would certainly be a distributed-computing hobbyist market for it if growing your farm meant adding a PCI card instead of a whole new box, and with negligible power consumption.