Internet proof of R109297 feasibility study

[paulunderwood]

OpenMPI version 5 has fault tolerance in the guise of User-Level Fault Mitigation (ULFM).

With about 100,000 CPU-core days needed for the certification of Rep-unit R109297 primality, I am wondering if it is feasible to run this over the Internet.

Would this mean that all clients have to use OpenMPI 5?

From the client's point of view, I guess ssh access is needed to an account for the MPI work.

How to handle changing IP addresses? Is there some kind of dynamic system for the server to handle the clients?

[bur]

Are you actively working on this? Would be an interesting project both in proving the first > 100,000 digits prime and, maybe even more importantly, in showing that distributed computing is possible for ECPP to some extent.

[paulunderwood]

Originally Posted by bur

Are you actively working on this? Would be an interesting project both in proving the first > 100,000 digits prime and, maybe even more importantly, in showing that distributed computing is possible for ECPP to some extent.

I suppose if anyone has a static IP with at least fast cores to keep up with my Threadripper. Static IP addresses is favoured, unless you know how to set up DDNS. Having installed ecpp-mpi plus GWNUM all that is needed by me is ssh access to an account on the client machine. The more cores a client machine has, the better. Some experimentation would be needed -- so let's start with a reasonably small PRP.

[kruoli]

If dynamic DNS can be used, I could provide you with some server for a test PRP. But isn't that reverse? I would think that the server (main MPI thread) needs the clients' IPs?

[paulunderwood]

Originally Posted by kruoli

If dynamic DNS can be used, I could provide you with some server for a test PRP. But isn't that reverse? I would think that the server (main MPI thread) needs the clients' IPs?

I will check, but I pretty sure DDNS would work. And yes, the central machine needs a host file that specifies the address and the number of slots (cores).

Thanks for being the first volunteer. Is there anyone else willing to participate?

[paulunderwood]

Dynamic DNS is perfectly usable. For those who do not know how it works: you install a program that periodically says to the DDNS "Hey, here is my current IP address". From my point of view I ssh via the DDNS address held by that server. DDNS can be had for free.

My master Threadripper has an i3 9100 as a slave over my LAN. The i3 slave does not stall the Threadripper.

With respect to fault tolerance, I do not know whether it will work well with ecpp-mpi and so will be initially using the standard MPI 4.

If things snowball I suggest we create a Whatsapp group to keep things moving along.

If there is any aspect of getting ecpp-mpi/GWNUM/ssh/DDNS to work that is problematic for you I can help.

[pinhodecarlos]

Paul, do you think the server can be hosted in the USA?

[paulunderwood]

Originally Posted by pinhodecarlos

Paul, do you think the server can be hosted in the USA?

Let me clear up a few things: The "master" "client" is being served by some "slave" "servers".

The master can be hosted anywhere. The limiting factor is the net. I currently have 20Mbs up and 70Mbs down.

[pinhodecarlos]

Last time I had someone helping me on CADO sieving the limitation on their side was the master IP server location, hence my question.
I'm fine for bandwidth speed (500Mbps D and U).

[paulunderwood]

Originally Posted by pinhodecarlos

Last time I had someone helping me on CADO sieving the limitation on their side was the master IP server location, hence my question.
I'm fine for bandwidth speed (500Mbps D and U).

I don't think bandwidth or latency are problems for ecpp-mpi over The Net. The data transferred is back and forth are small and worker jobs are relatively lengthy.

[pinhodecarlos]

So why have you said the limiting factor was the net.