climateprediction.net (CPDN) home page
Thread 'processors, memory, performance and heat.'

Thread 'processors, memory, performance and heat.'

Message boards : Number crunching : processors, memory, performance and heat.
Message board moderation

To post messages, you must log in.

Previous · 1 · 2 · 3 · Next

AuthorMessage
ProfileDave Jackson
Volunteer moderator

Send message
Joined: 15 May 09
Posts: 4538
Credit: 19,004,017
RAC: 21,574
Message 70855 - Posted: 21 Apr 2024, 11:50:20 UTC

Finally gotten around to ordering an extra 64GB of RAM. I have ordered PC3200 RAM while the 32GB I have is PC2800. Will I be better off ditching the old and just running with 64GB to take advantage of the slightly higher speed or keeping it to have 96GB total?

Main reason for upgrading is to increase the maximum number of OIFS tasks I can run at once.
ID: 70855 · Report as offensive     Reply Quote
ProfileAlan K

Send message
Joined: 22 Feb 06
Posts: 491
Credit: 30,975,898
RAC: 14,500
Message 70856 - Posted: 21 Apr 2024, 22:19:24 UTC - in response to Message 70855.  

My guess is that it really depends on how many cores your processor has and therefore how many tasks it can run at any one time. I am not sure that having differeent speed memory chips makes much difference to perforamnce.
ID: 70856 · Report as offensive     Reply Quote
Jean-David Beyer

Send message
Joined: 5 Aug 04
Posts: 1120
Credit: 17,202,915
RAC: 2,154
Message 70857 - Posted: 21 Apr 2024, 23:56:15 UTC - in response to Message 70856.  

My guess is that it really depends on how many cores your processor has and therefore how many tasks it can run at any one time. I am not sure that having differeent speed memory chips makes much difference to perforamnce.


It also depends on how big the cache on the processor chip(s) is.
ID: 70857 · Report as offensive     Reply Quote
ProfileDave Jackson
Volunteer moderator

Send message
Joined: 15 May 09
Posts: 4538
Credit: 19,004,017
RAC: 21,574
Message 70858 - Posted: 22 Apr 2024, 6:42:04 UTC

My guess is that it really depends on how many cores your processor has and therefore how many tasks it can run at any one time.

AMD Ryzen 7 3700X 8-Core Processor [Family 23 Model 113 Stepping 0] When it arrives, I might try with just the 64GB of faster ram and then tweak the motherboard settings to take advantage of it. I never to run more than 7 tasks at a time so that still leaves a reasonable bit of headroom. If I don't see any improvement, I might put the other 32 back in and have a ram disk.
ID: 70858 · Report as offensive     Reply Quote
SolarSyonyk

Send message
Joined: 7 Sep 16
Posts: 262
Credit: 34,915,412
RAC: 16,463
Message 70859 - Posted: 22 Apr 2024, 16:07:22 UTC

The CPDN tasks seem to really hammer on memory, so I'd expect, if it's sufficient to run the task quantities your chip can reasonably support, "less faster memory" would be better than "more slower memory."
ID: 70859 · Report as offensive     Reply Quote
ProfileDave Jackson
Volunteer moderator

Send message
Joined: 15 May 09
Posts: 4538
Credit: 19,004,017
RAC: 21,574
Message 70860 - Posted: 22 Apr 2024, 16:28:11 UTC - in response to Message 70859.  

The CPDN tasks seem to really hammer on memory, so I'd expect, if it's sufficient to run the task quantities your chip can reasonably support, "less faster memory" would be better than "more slower memory."

I think 64GB for an 8 core CPU is probably sufficient. Some tasks in testing have gone up above 9GB peak per task recently but they have the BOINC debugging code enabled which adds a lot to the overhead. If at some point the demand goes up even more, I can either go up to the max on the motherboard of 128GB or run fewer tasks. Currently I am running one of the memory intensive tasks and six WAH2ri tasks and free tells me that swap is unused. I do wonder if once I double my RAM I could get rid of swap altogether?
ID: 70860 · Report as offensive     Reply Quote
Jean-David Beyer

Send message
Joined: 5 Aug 04
Posts: 1120
Credit: 17,202,915
RAC: 2,154
Message 70861 - Posted: 22 Apr 2024, 20:31:47 UTC - in response to Message 70860.  

The CPDN tasks seem to really hammer on memory, so I'd expect, if it's sufficient to run the task quantities your chip can reasonably support, "less faster memory" would be better than "more slower memory."

I think 64GB for an 8 core CPU is probably sufficient. Some tasks in testing have gone up above 9GB peak per task recently but they have the BOINC debugging code enabled which adds a lot to the overhead. If at some point the demand goes up even more, I can either go up to the max on the motherboard of 128GB or run fewer tasks. Currently I am running one of the memory intensive tasks and six WAH2ri tasks and free tells me that swap is unused. I do wonder if once I double my RAM I could get rid of swap altogether?

In the very old days, it was recommended to have twice as much disk space for swap as the machine had RAM. Later, as RAM cost less and people put more and more RAM in their machines, this recommended amount was reduced to the same amount of swap as RAM. But RAM has gotten so large that I run my current machine with the same amount od swap as it had when I bought it. It had 32 GBytes of RAM when I bought it. though it has 128 GBytes now. It runs Linux. It has been running 24/7 for about a week and has used about 15 Megabytes of RAM: not much. It is running 11 Boinc processes at the moment.

$ free -hw
              total        used        free      shared     buffers       cache   available
Mem:          125Gi       6.1Gi       1.7Gi       1.5Gi       764Mi       116Gi       116Gi
Swap:          15Gi        15Mi        15Gi

ID: 70861 · Report as offensive     Reply Quote
ProfileDave Jackson
Volunteer moderator

Send message
Joined: 15 May 09
Posts: 4538
Credit: 19,004,017
RAC: 21,574
Message 70868 - Posted: 25 Apr 2024, 13:35:13 UTC

The decision turns out to be academic. The old Corsair memory is fat enough that it won't fit in the slots alongside the new memory. However if I ever decide 64GB is not enough two more of the crucial modules I just bought will fit enabling me to double again.
ID: 70868 · Report as offensive     Reply Quote
wateroakley

Send message
Joined: 6 Aug 04
Posts: 195
Credit: 28,332,519
RAC: 10,361
Message 70874 - Posted: 27 Apr 2024, 18:05:52 UTC - in response to Message 70860.  

I do wonder if once I double my RAM I could get rid of swap altogether?
Probably not a good idea to get rid of the swap file. The OS behaviour would be unpredictable.
ID: 70874 · Report as offensive     Reply Quote
ProfileDave Jackson
Volunteer moderator

Send message
Joined: 15 May 09
Posts: 4538
Credit: 19,004,017
RAC: 21,574
Message 70875 - Posted: 28 Apr 2024, 4:54:44 UTC - in response to Message 70874.  

I do wonder if once I double my RAM I could get rid of swap altogether?
Probably not a good idea to get rid of the swap file. The OS behaviour would be unpredictable.
Perhaps though swap hasn't been touched since I did the upgrade. I might look at substantially reducing the size from the current 64GB the next time I do a reinstall from scratch. Though that might be a while!
ID: 70875 · Report as offensive     Reply Quote
Jean-David Beyer

Send message
Joined: 5 Aug 04
Posts: 1120
Credit: 17,202,915
RAC: 2,154
Message 70876 - Posted: 28 Apr 2024, 11:44:29 UTC - in response to Message 70875.  
Last modified: 28 Apr 2024, 11:46:37 UTC

Perhaps though swap hasn't been touched since I did the upgrade. I might look at substantially reducing the size from the current 64GB the next time I do a reinstall from scratch. Though that might be a while!


I would keep a swap file. My machine has a swap file that is 16 GBytes and it is using 1/4 GByte of swap. If I boot the machine, it goes a couple of days without using any swap, but it slowly goes up a little bit. It happens when I am asleep, so it probably has something to do with writing backups to external hard drive. Machine mostly runs Boinc 24/7, but I do some other things during the day. It has done no CPDN since about last June because there is no Linux work from CPDN.

My machine has 128 GBytes of RAM, yet, in time, I use a little swap. This machine has been up almost two weeks and it has used a trifle of swap.
top - 07:32:33 up 13 days, 21:41,  2 users,  load average: 13.37, 13.58, 13.40
Tasks: 476 total,  17 running, 459 sleeping,   0 stopped,   0 zombie
%Cpu(s):  0.6 us,  0.3 sy, 80.6 ni, 18.3 id,  0.0 wa,  0.2 hi,  0.0 si,  0.0 st
MiB Mem : 128074.0 total,   1635.3 free,   6490.1 used, 119948.7 buff/cache
MiB Swap:  15992.0 total,  15717.7 free,    274.2 used. 117948.7 avail Mem 



$ free -hw
              total        used        free      shared     buffers       cache   available
Mem:          125Gi       6.3Gi       1.6Gi       2.4Gi       1.1Gi       116Gi       115Gi
Swap:          15Gi       274Mi        15Gi

ID: 70876 · Report as offensive     Reply Quote
Glenn Carver

Send message
Joined: 29 Oct 17
Posts: 1049
Credit: 16,432,494
RAC: 17,331
Message 70878 - Posted: 28 Apr 2024, 12:38:59 UTC - in response to Message 70875.  
Last modified: 28 Apr 2024, 12:39:05 UTC

A practical reason to have swap these days is if you intend using 'hibernate', which dumps the memory to swap space on the drive (i.e. configure swap to be 1.5x installed RAM). Otherwise, modern OSes work fine without swap as long as you have sufficient memory for what it is you intend to do with the machine.
ID: 70878 · Report as offensive     Reply Quote
Jean-David Beyer

Send message
Joined: 5 Aug 04
Posts: 1120
Credit: 17,202,915
RAC: 2,154
Message 70879 - Posted: 28 Apr 2024, 17:40:13 UTC - in response to Message 70878.  

A practical reason to have swap these days is if you intend using 'hibernate', which dumps the memory to swap space on the drive (i.e. configure swap to be 1.5x installed RAM).

I do not even know how to make my Linux machine hibernate. So I never do it.

As delivered, it had 32 GBytes of RAM and 16 Gbytes of swap. I upped the RAM to 128 GBytes, but saw no reason to increase the size of the swap area of my solid state drive. No reason to reduce it either. My Boinc data are on a partition of its own on a spinning 4 TByte hard drive, I cannot imagine having 192 GBytes of swap. Running two weeks 24/7, my machine is currently using 278.8 Megabytes of swap while running 13 Boinc tasks. That is hardly any.

Otherwise, modern OSes work fine without swap as long as you have sufficient memory for what it is you intend to do with the machine.


If It did not have swap space, where would the OS have put the 278.8 Megabytes of stuff?
ID: 70879 · Report as offensive     Reply Quote
SolarSyonyk

Send message
Joined: 7 Sep 16
Posts: 262
Credit: 34,915,412
RAC: 16,463
Message 70883 - Posted: 29 Apr 2024, 18:21:05 UTC - in response to Message 70874.  

Probably not a good idea to get rid of the swap file. The OS behaviour would be unpredictable.


Any modern operating system runs just fine with no swapfile, and has for a couple decades now. Just don't run it out of RAM, unless you want the OOM reaper to show up.
ID: 70883 · Report as offensive     Reply Quote
noderaser
Avatar

Send message
Joined: 17 Nov 04
Posts: 16
Credit: 1,938,125
RAC: 2,007
Message 70905 - Posted: 6 May 2024, 17:44:04 UTC

If you are running on Windows, TThrottle is a tool you can use to scale your BOINC computation on the fly to hold a specified CPU temperature. I make use of it on all my Windows machines, particularly my laptop which gets quite toasty if I let it run at 100% unchecked.
Click Here to see My Detailed BOINC Stats
ID: 70905 · Report as offensive     Reply Quote
ProfileDave Jackson
Volunteer moderator

Send message
Joined: 15 May 09
Posts: 4538
Credit: 19,004,017
RAC: 21,574
Message 70934 - Posted: 7 Jun 2024, 14:36:41 UTC

My Ryzen7 is crashing too often to be thought of as reliable. Will shortly be replaced by a 9 (12 real cores instead of 8) and 64GB RAM. I will probably use a maximum of 8 cores for crunching at a time. Hopefully it will be more reliable as well as able to return tasks a little faster.
ID: 70934 · Report as offensive     Reply Quote
Jean-David Beyer

Send message
Joined: 5 Aug 04
Posts: 1120
Credit: 17,202,915
RAC: 2,154
Message 70963 - Posted: 8 Jun 2024, 14:18:41 UTC - in response to Message 70622.  
Last modified: 8 Jun 2024, 14:19:36 UTC

I now have three CPDN tasks running on my Linux machine and I get this:

# date; perf stat -e cache-references,cache-misses,cycles,instructions,branches,faults,migrations
Sat Jun  8 10:07:00 EDT 2024
^C
 Performance counter stats for 'system wide':

    18,187,914,185      cache-references                                            
     6,183,637,575      cache-misses              #   33.999 % of all cache refs    
 3,558,923,737,838      cycles                                                      
 4,815,823,664,127      instructions              #    1.35  insn per cycle         
   683,006,914,699      branches                                                    
         1,061,496      faults                                                      
            15,715      migrations                                                  

      60.983597747 seconds time elapsed


So about 67% of memory references are obtained from the cache. My processor has a fairly large cache.

CPU type 	GenuineIntel
Intel(R) Xeon(R) W-2245 CPU @ 3.90GHz [Family 6 Model 85 Stepping 7]
Number of processors 	16

Operating System 	Linux Red Hat Enterprise Linux
Red Hat Enterprise Linux 8.10 (Ootpa) [4.18.0-553.5.1.el8_10.x86_64|libc 2.28]
BOINC version 	7.20.2
Memory 	125.08 GB
Cache 	16896 KB

ID: 70963 · Report as offensive     Reply Quote
ProfileDave Jackson
Volunteer moderator

Send message
Joined: 15 May 09
Posts: 4538
Credit: 19,004,017
RAC: 21,574
Message 71400 - Posted: 7 Sep 2024, 9:17:04 UTC

Been running tasks under WINE since the current batches started. I notice all but 4 of the top 20 computers are Ryzen ones. I see I have now made the top 20, mostly running 12 out of 16 real cores which is the most I can run without the temperature spikes going above 90C. Maybe I should have looked at a cooling system capable of dissipating more heat? I expect to be able to increase the number of cores in action when the weather cools down further.
ID: 71400 · Report as offensive     Reply Quote
Glenn Carver

Send message
Joined: 29 Oct 17
Posts: 1049
Credit: 16,432,494
RAC: 17,331
Message 71401 - Posted: 7 Sep 2024, 9:41:55 UTC - in response to Message 71400.  
Last modified: 7 Sep 2024, 9:46:14 UTC

Probably because there's a lot of gamers on the CPDN books :). AMD has never been the best for number crunching. Intel equivalent chips have better single core performance, better multi-core performance, more cores, and pertinent to your comment a higher max Tjunction limit. Interesting the latest AMD offering have finally caught up on single core performance but not multi-core. Leaks for Arrow Lake performance are looking good though.

Definitely improve the cooling if you hit 90C without all the cores running.
ID: 71401 · Report as offensive     Reply Quote
Richard Haselgrove

Send message
Joined: 1 Jan 07
Posts: 1061
Credit: 36,700,823
RAC: 9,977
Message 71402 - Posted: 7 Sep 2024, 10:50:19 UTC - in response to Message 71401.  

There's a worrying article about recent Intel chips in this month's UK edition of PC PRO magazine (no. 361, cover date October 2024). It says that "... There have long been reports of instability with Intel's "Raptor Lake" processors, with systems crashing under heavy load, and even dead chips." Apparently, it appears to mainly affect the bigger i7 and i9 versions

Reports of problems with the Intel processors have been swirling since late 2022, particularly among the overclocking community, who are most likely to trigger the potentially damaging voltages.

Intel had remained tight-lipped on the issue until late July, when it finally admitted that 13th and 14th generation desktop processors with 65W or higher base power "could be affected by elevated voltage issues". Worse, chips could be irreparably damaged by the excess voltage.

It followed up with the microcode patch in mid-August, but as this patch can't be distributed via Windows Update, it's reliant on either PC manufacturers contacting affected customers or buyers seeking out the BIOS update by themselves.
The electronic edition of the full article seems to be behind a paywall, but I have a full scan available on request.
ID: 71402 · Report as offensive     Reply Quote
Previous · 1 · 2 · 3 · Next

Message boards : Number crunching : processors, memory, performance and heat.

©2024 cpdn.org