Config file changes to consolidate workers - moving to a 22-core config

NookieN · 2017-07-29, 18:16

Hopefully this is a ridiculously simple question. I want to go from 3 workers down to one and keep my current assignments and progress saved for them. Can I just change WorkerThreads to 1 in local.txt and move all of my assignments to Worker 1 in worktodo.txt? Will that confuse anything on primenet?

Background: I have been running a 5820K with 3 workers, 2 threads each. I recently got my hands on a 22-core Xeon and I'm playing around with that. It looks like it would be too memory-bound to run more than one worker, but with 22 threads in one worker it could cut through a first LL test in the current range in just over a day.

thyw · 2017-07-29, 19:41

Yes, it should work, i have never had a problem with this - but I still advice you to back up them in the rare case it messes up.

GP2 · 2017-07-29, 22:20

It's not just the WorkerThreads parameter you need to adjust.

You should also modify the CoresPerTest line (in version 29) or ThreadsPerTest line (in versions 28 and earlier), or create this line if it doesn't exist already.

Note, are all 22 cores on the same processor, or do you have dual processors with 11 cores each? If the latter, it would be more efficient to run WorkerThreads=2 and CoresPerTest=11 rather than WorkerThreads=1 and CoresPerTest=22. Which model of chip do you have?

If you have been running the very latest (not yet fully officially released) version 29.2, then you should also delete the gwnum.txt file, if it exists, when you make the above changes to local.txt, because all your timing information will be different going forward.

NookieN · 2017-07-29, 23:02

Thanks for the replies. I'm using v28.9b2 right now. I should probably go ahead and benchmark v29.2 for comparison. The chip I'm running is a single 22-core chip, Broadwell-E architecture, very similar to the 2699v4 though there are some slight differences:

http://ark.intel.com/products/91317/...Cache-2_20-GHz

The memory controller is limited to DDR-2400 and although I can mess with that in the EFI settings I haven't found a stable boot configuration yet (nor do I really expect to). If nothing else I might push the memory timings a bit.

Madpoo · 2017-07-31, 15:19

Quote:

Originally Posted by NookieN

Thanks for the replies. I'm using v28.9b2 right now. I should probably go ahead and benchmark v29.2 for comparison. The chip I'm running is a single 22-core chip, Broadwell-E architecture, very similar to the 2699v4 though there are some slight differences:

http://ark.intel.com/products/91317/...Cache-2_20-GHz

The memory controller is limited to DDR-2400 and although I can mess with that in the EFI settings I haven't found a stable boot configuration yet (nor do I really expect to). If nothing else I might push the memory timings a bit.

You're probably better off going with version 29.2 since it's easier to set the cores that worker will use (if you wanted to fuss with it at all, that is). On a single-CPU system like yours, it's probably fine to just let 'er rip, but on my dual CPU systems I still found, even with v29, that it's better to manually tell it which cores to use for which worker, otherwise it assigns cores from different chips to the same worker or just generally didn't seem to work out as well.

As for consolidating, yeah, if you move all of your work to the [Worker #1] section of your worktodo.txt file that'll do just fine. At worst I could imagine the software thinking you have too much work queued up and it could unreserve some until it gets the right rolling average figured out for your new system. That's not likely since you're also switching from a lower core-count system to the 22-core... if anything it seems like it would get more assignments since it can kick through them faster.

Anyway, it'll all sort itself out in the end, and if you have problems, come back here or PM me if you lose any assignments or whatever.

vsuite · 2017-08-02, 01:06

I've found that it's best to first disconnect the internet on the computer I'm making configuration changes to. This way, prime95 does not re-configure changes and write them to the server. Then.

1 Exit (not close) prime95.
2 Copy the new prime95.exe over your current installation.
3 Restart prime95
4 Re-configure your Workers
5 Exit prime95
6 Edit the worktodo.txt file to tidy things up
7 Restart prime95
8 Reconnect the internet

NookieN · 2017-08-10, 05:43

Well the worker consolidation went well. After moving 3 75M assignments to worker 1 it chewed through them in about 2.5 days. It turns out that v29.x is very much necessary for this chip. I received two double-check assignments in the 45M range and noticed they were taking about 1.75ms/itr on all 22 cores vs around 1.46ms/itr for the 75M exponents.

Looking at my v28 benchmarks there was a donut hole phenomenon where exponents between 1536k and 3072k were taking longer per itr than larger exponents. Even within a specific FFT size (e.g. 4480k) I saw 1.76ms on one exponent and 2.59ms on another. v29 does fix that and I have an 81M test cooking along at 1.58ms/itr at the moment.

I can post all the benchmarks to the perpetual thread if anyone is interested. Another fun fact about this Xeon is that while it is rated at 2.2GHz it likes to run at 2.6GHz most of the time even with all cores active.

GP2 · 2017-08-10, 17:57

Quote:

Originally Posted by NookieN

It turns out that v29.x is very much necessary for this chip.

The software download page still offers only 28.10

This might soon become an issue as both AMD and Intel start selling chips with large numbers of cores at mass consumer-friendly prices.

2017-07-29, 19:41	#2
thyw Feb 2016 ! North_America 83 Posts	Yes, it should work, i have never had a problem with this - but I still advice you to back up them in the rare case it messes up. Last fiddled with by thyw on 2017-07-29 at 19:42

2017-07-29, 22:20	#3
GP2 Sep 2003 5·11·47 Posts	It's not just the WorkerThreads parameter you need to adjust. You should also modify the CoresPerTest line (in version 29) or ThreadsPerTest line (in versions 28 and earlier), or create this line if it doesn't exist already. Note, are all 22 cores on the same processor, or do you have dual processors with 11 cores each? If the latter, it would be more efficient to run WorkerThreads=2 and CoresPerTest=11 rather than WorkerThreads=1 and CoresPerTest=22. Which model of chip do you have? If you have been running the very latest (not yet fully officially released) version 29.2, then you should also delete the gwnum.txt file, if it exists, when you make the above changes to local.txt, because all your timing information will be different going forward. Last fiddled with by GP2 on 2017-07-29 at 22:22

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
how to config prime95 for my PC...plz help	Flexagon	Information & Answers	7	2018-03-21 11:05
prime95 config file changes	bgbeuning	Software	8	2018-02-11 16:58
CudaLucas correct config	jpalo	GPU Computing	8	2017-08-06 15:35
Can an OS config be OK after RAM failed?	RickC	Hardware	8	2010-10-28 03:31
Startup config for Quad Core on Linux	dswanson	Hardware	4	2008-01-28 16:29

2017-07-29, 18:16	#1
NookieN Aug 2002 3A₁₆ Posts	Config file changes to consolidate workers - moving to a 22-core config Hopefully this is a ridiculously simple question. I want to go from 3 workers down to one and keep my current assignments and progress saved for them. Can I just change WorkerThreads to 1 in local.txt and move all of my assignments to Worker 1 in worktodo.txt? Will that confuse anything on primenet? Background: I have been running a 5820K with 3 workers, 2 threads each. I recently got my hands on a 22-core Xeon and I'm playing around with that. It looks like it would be too memory-bound to run more than one worker, but with 22 threads in one worker it could cut through a first LL test in the current range in just over a day. Last fiddled with by NookieN on 2017-07-29 at 18:18 Reason: Clarified title for HW forum

2017-07-29, 23:02	#4
NookieN Aug 2002 2·29 Posts	Thanks for the replies. I'm using v28.9b2 right now. I should probably go ahead and benchmark v29.2 for comparison. The chip I'm running is a single 22-core chip, Broadwell-E architecture, very similar to the 2699v4 though there are some slight differences: http://ark.intel.com/products/91317/...Cache-2_20-GHz The memory controller is limited to DDR-2400 and although I can mess with that in the EFI settings I haven't found a stable boot configuration yet (nor do I really expect to). If nothing else I might push the memory timings a bit.

2017-08-02, 01:06	#6
vsuite Jan 2010 2×3×19 Posts	I've found that it's best to first disconnect the internet on the computer I'm making configuration changes to. This way, prime95 does not re-configure changes and write them to the server. Then. 1 Exit (not close) prime95. 2 Copy the new prime95.exe over your current installation. 3 Restart prime95 4 Re-configure your Workers 5 Exit prime95 6 Edit the worktodo.txt file to tidy things up 7 Restart prime95 8 Reconnect the internet

2017-08-10, 05:43	#7
NookieN Aug 2002 3A₁₆ Posts	Well the worker consolidation went well. After moving 3 75M assignments to worker 1 it chewed through them in about 2.5 days. It turns out that v29.x is very much necessary for this chip. I received two double-check assignments in the 45M range and noticed they were taking about 1.75ms/itr on all 22 cores vs around 1.46ms/itr for the 75M exponents. Looking at my v28 benchmarks there was a donut hole phenomenon where exponents between 1536k and 3072k were taking longer per itr than larger exponents. Even within a specific FFT size (e.g. 4480k) I saw 1.76ms on one exponent and 2.59ms on another. v29 does fix that and I have an 81M test cooking along at 1.58ms/itr at the moment. I can post all the benchmarks to the perpetual thread if anyone is interested. Another fun fact about this Xeon is that while it is rated at 2.2GHz it likes to run at 2.6GHz most of the time even with all cores active.