View Single Post
Old 2021-08-22, 00:37   #8
kriesel's Avatar
Mar 2017
US midwest

19×311 Posts

I think you have some choices. (Some of this is already covered somewhat by previous posts made by others while I was writing this)

Assuming 1GB ram, 8GB SD storage, Linux, mprime, & wifi equipped (802.11b/g/n):

A) mprime & PRP / GEC at low proof power is better than none. Run a low proof power to reduce storage space required for proof generation.
At power 7, that's ~3.2GB temp space, 400MB proof file, for p=100M, proportional to exponent (so ~3.5GB & 440MB at p=110M) both of which can get cleared automatically after proof is uploaded to the server. Still saves ~99.2% of DC effort.
At power 6, that's 1.6GB temp space, ~800MB proof file, and still saves ~98.5% of DC effort. (~1.76GB & 880MB =2.64GB at p=110M)
Minimum proof power in prime95/mprime is 5; 800MB temp space, & still saves ~97% of DC effort.
(George please correct any wrong numbers given)
How many GB of storage does your minimum install use up?

B) use other options to limit mprime's local footprint.
Put the executable, the temp file or proof file or all on a LAN server, and perhaps the more actively used working files stay on the stick's working directory.
Use Wifi or other connectivity options to access server storage. Use the separate settings for "optional directory to hold large temporary files" and "optional directory to hold archived proofs"
And as George said, reduce mprime number of backup files. Maybe you already have.

C) Use Windows which can support bigger SD cards (32GB) as local storage (but maybe that requires swapping out the Intel compute sticks or more cost than tolerable)

D) (Not recommended due to low efficiency, ~50%, sort of a waste of time compared to PRP/GEC/proof-gen once) Run PRP/GEC without proof, and PRP DC your own PRP first tests or partner with another contributor to DC each other's.
Recent Mlucas, or mprime V29.8 don't implement proof generation, so could be run on hardware that lacks space for proof gen and also lacks access to network shares with sufficient space. No proof generation means a full PRP DC is required.

E) Sell your idled 12 sticks, and buy newer hardware with adequate capacity for today's workloads.
I haven't used them, but some like Raspberry Pi. Recent Raspberry Pis have gigabit wired ethernet jacks.
Those could probably be network booted and use server hosted working directories, so make use of TB space available for sharing, temp space, proof file, backup files, etc. They're ARM based so would involve using Mlucas.
Mlucas does not have PRP proof file generation, yet.
Another way to go is a modest priced AMD GPU which could do the equivalent of 12-20 sticks of computing or more depending on models, in one unit to manage.
(For scale, 1 RadeonVII GPU can do about a wavefront PRP a day. An RX480 in ~5 days; RX550 ~18 days.)

F) Run some P-1 factoring. There's always a need for that at the wavefront. If I could do it with 1GB ram in a Windows laptop, it should be feasible in 1GB ram in Linux.

G) (Not recommended because it's very inefficient and less reliable)
Convert legitimate PRP first test assignments into bootleg LL first test,
settle for the 50% detection rate of errors that the Jacobi symbol check provides,
and stick the project with 102% or higher DC overhead.

Have fun!

Last fiddled with by kriesel on 2021-08-22 at 01:24
kriesel is online now   Reply With Quote