20200416, 07:36  #1 
"Simon Josefsson"
Jan 2020
Stockholm
41_{8} Posts 
Best use of highmemory machines?
Hi. I have a couple of machines with larger amounts of memory that I would like to use for the best purpose for GIMPS (with a slight personal bias towards actually making history and find a new prime). The machines have dual Xeon E5 CPUs with 812 cores each, such as 2x2673v3, 2x2640v2, 2x2670v1 and have 128GB or 256GB now but could be rearranged into fewer machines but with 512GB or even 1TB of RAM.
I'm doing PRP's now but that doesn't really utilize a lot of the memory. What are the other options? It looks like P1 factoring can make use of the memory, but I don't know how to benchmark this in a good way, can someone help? Is there a way to benchmark various parameters such as: 1) Number of CPU cores 124 3) Number of workers 2) Amount of memory 32GB, 64GB, 96GB, 128GB, 196GB, 256GB Currently I am doing some P1 factoring on my machines, but they are doing PRP's at the same time so it is hard to tell if the P1 worker is working optimally. Thanks, Simon 
20200416, 13:47  #2 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest
2×3×5×7×29 Posts 
P1 is the most memoryhungry of the TF, P1, primality test sequence of working toward finding the next Mersenne prime by elimination. Up to at least 32GB per P1 worker is useful. There are many exponents that have not had P1 factoring applied, or only to very low limits. It's best to go for the full limits initially, not to low limits first as some users do. I suspect applying an entire Xeon to one P1 worker would be best cache usage and performance. Test for yourself. For the B1 and B2 bounds, I recommend using the PrimeNet values on James Heinrich's site for the exponent;
https://www.mersenne.ca/exponent/332233123 for example. The approach of testing to full limits immediately, without any other smallerbounds runs, is what RDS has supported. See https://www.mersenneforum.org/showpo...9&postcount=20 for test cases performed in gpuowl. These empirically confirmed the RDS position. Also possibly useful or interesting background are https://www.mersenneforum.org/showpo...7&postcount=12 https://www.mersenneforum.org/showpo...4&postcount=17 https://www.mersenneforum.org/showpo...9&postcount=30 The best way to benchmark P1 on your hardware is to make actual runs. The probability of finding a factor can be calculated using https://www.mersenne.ca/prob.php I don't know whether homogeneous (all P1 on a system) or heterogeneous (mixing P1 and primality testing on one system) is better performance. With differing computation types there's need for a common basis of comparison. I usually use P1 runtime vs. P1factor odds x 2.04 LL tests or x 2 PRP tests. Please post your benchmark results. Last fiddled with by kriesel on 20200416 at 14:00 
20200416, 17:07  #3 
P90 years forever!
Aug 2002
Yeehaw, FL
1111001010101_{2} Posts 
Sorry to disappoint you, but quantity of RAM is not an important ingredient in finding new Mersenne primes.
Yes, P1 can use more RAM but it is only slightly beneficial. 
20200416, 17:54  #4 
"6800 descendent"
Feb 2005
Colorado
2^{4}·43 Posts 
If finding huge (really huge) factors sounds interesting, you can put all that memory to use using GNFS or SNFS to find factors for the Cunningham Project.

20200416, 18:13  #5 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest
2·3·5·7·29 Posts 
Having all memory channels populated improves benchmarks and performance, yes? With an existing system that's already populated with DIMMs of set size, that may mean having a lot more GB installed than is strictly necessary for the data size. For example, I don't need 16GB to run one worker, but it performs much better with two 8GB DIMMs than one on an i74790.
Last fiddled with by kriesel on 20200416 at 18:17 
20200416, 19:40  #6  
Nov 2003
2^{2}·5·373 Posts 
Quote:
linear algebra. There is a big backlog. Lots of memory does not really help GIMPS. 

20200418, 12:46  #7 
Bamboozled!
"๐บ๐๐ท๐ท๐ญ"
May 2003
Down not across
2×3×17×109 Posts 

Thread Tools  
Similar Threads  
Thread  Thread Starter  Forum  Replies  Last Post 
High Bandwidth Memory  tha  GPU Computing  4  20150731 00:21 
High weight k's  kar_bon  Riesel Prime Data Collecting (k*2^n1)  26  20130911 23:12 
"Hybrid Memory Cube" offers 1 Tb/s memory bandwith at just 1.4 mW/Gb/s  ixfd64  Hardware  4  20111214 21:24 
how high will CRUS go  MiniGeek  Conjectures 'R Us  1  20101108 20:50 
High IPC bad for GIMPS?  db597  Hardware  4  20050715 20:47 