Prime Gap Search latest version of the c code

[pinhodecarlos]

Here is the latest version of the c code (version 0.05.g), along with Windows executables and linux binaries.

Originally Posted by Antonio

Here is the latest version of the c code (version 0.05.g), along with Windows executables (now reduced to two versions - one for pre-Haswell the other for Haswell and later processors, as discussed in earlier posts).
There are only minor changes in this version:

// version 0.05.g    Fixed quick sort, added optimized bubble sort.            Antonio Key
//                    Replaced add30 & sub30 constant arrays with a single
//                    wheel30 constant array - they became identical in 0.05.
//                    Added ETA to status display line.
//                    Removed some of the old commented-out code.
//                    Modified the smart check on input data.

// version 0.05.f    next_prime, prec_prime tuning.                            Antonio Key

// version 0.05.e    Hensel lifting.                                            Robert Gerbicz

// version 0.05.d    Fermat & Euler-Plumb PRP tests using Montgomery math.    Dana Jacobsen

// version 0.05        Bug fix.                                                Robert Gerbicz

The compiled versions use the optimized bubble sort, I don't think there is any advantage to using the quick sort for the small number of items we have to order at the end of each stage and the bubble sort may be faster.
The modification to the input data smart check came about as a result of my further investigation into tuning the code. I have found that for my i5-3570k running 4 threads I can increase the throughput to 27.5e9 n/sec using bs= 18, the old smart check would issue a warning for bs > 16, which didn't seem reasonable. The smart check now only displays a warning if sb <16 or (sb - bs) < 3.
My input parameters have now changed from (ignore the n1 and n2 parameters - which are my test range values only):

gap5_f_ivybridge  -n1 4248785e12 -n2 4248789e12 -gap 1250 -delta 150 -sb 24 -bs 16 -t 4  -mem 12
throughput 26.4e9 n/sec.

to:

gap5_g -n1 4248785e12 -n2 4248789e12 -gap 1250 -delta 155 -sb 24 -bs 18 -t 4 -mem 12
throughput 27.5e9 n/sec.

Antonio and Thomas, please use this thread so people can easily follow up the latest binaries.

[pinhodecarlos]

What's the relation between the client speed and memory used? Let´s imagine I have a bunch of cores with only 1 to 2GB of memory available per core. Is it worth to have for example a client that uses 500MB memory/core?

[ET_]

Originally Posted by pinhodecarlos

What's the relation between the client speed and memory used? Let´s imagine I have a bunch of cores with only 1 to 2GB of memory available per core. Is it worth to have for example a client that uses 500MB memory/core?

I asked the same question to run on an Odroid-C2 (2GB). The answer (with the related command-line and source modifications) should be buried inside this thread.

I will search for it, be back soon.

P.S. Here it is: Here starting from post #70.

[pinhodecarlos]

Thank you Luigi. Did try to run on my wife's laptop but the client needs at least 3GB. Also I am thinking on something else....

[ET_]

Originally Posted by pinhodecarlos

Thank you Luigi. Did try to run on my wife's laptop but the client needs at least 3GB. Also I am thinking on something else....

I was running gap on my i5: 2 threads, 2.1GB (of 4GB present on the PC), so I assume you can run it safely on the laptop...

Wait... you are on Windows, aren't you?

[pinhodecarlos]

Yes, I am on windows.

[danaj]

I was able to get it running with 3 threads on Windows in about 2GB. I used "-sb 23 -bs 14", and lowered MAX_SIZE to 1LL<<26. The latter didn't have nearly the memory impact I thought it would from the comments (lowering it further seemed to have basically no memory effect).

I wrote some debugging printfs to help figure out where memory was going ... Perhaps I should write those up in a more friendly one-line output. As we progress higher in the search the memory use will get worse.

Unfortunately the PC has 4GB and for some reason that gives it under 3GB actually usable with Windows 10 64-bit. Task Viewer says it has about 600MB free when running the gap program, and firing up Firefox or Emby is much slower.

[Antonio]

The (Windows) version of code attached requires -mem (1.55 + 0.05t) to run, where t is the number of threads.
To test it I used the command line:- gap5_size -n1 5501904e12 -n2 5501907e12 -gap 1340 -delta 150 -sb 23 -bs 14 -t 1 -mem 1.6
It is about 1/2 the speed of the 0.05_g code on my machine, although I didn't try re-tuning delta, which may help.

[pinhodecarlos]

Was wondering if I will be able to run the client on a MINIX NEO Z83-4.

[ET_]

Originally Posted by pinhodecarlos

Was wondering if I will be able to run the client on a MINIX NEO Z83-4.

The code has a couple of function hand-tuned in x86 assembler.
They don't compile on ARM machines.

[pinhodecarlos]

Originally Posted by ET_

The code has a couple of function hand-tuned in x86 assembler.
They don't compile on ARM machines.

This one has an Intel processor.