PrimeGrid Races 2018

Page 10 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

StefanR5R

Elite Member
Dec 10, 2016
4,628
5,572
136
So, just like in the previous race, my single-socket Xeon E5 (Linux Mint) is quite a bit slower than the dual-socket E5s (Gentoo and OpenSUSE). But this time the single-socket E5 isn't running any GPU jobs on the side. The primegrid_llr processes show up with less %CPU in "top" on the 1P host compared to those on the 2Ps, for no apparent reason.

On a whim, I tried this:
$ sudo /etc/init.d/lightdm stop
And suddenly, %CPU went up to the level that I am seeing on the 2Ps.
 

biodoc

Diamond Member
Dec 29, 2005
6,045
1,915
136
That's disturbing. I'm assuming that's Mint 18.3. Have you done all the system updates/upgrades and are running the latest kernel? I have the latest kernel version for 18.3 which is 4.13.0-38-generic. I'm running a multithreaded Quantum chem app with 4 logical cores and it is spot on with 400% processor usage according to top.
 

StefanR5R

Elite Member
Dec 10, 2016
4,628
5,572
136
I spoke too soon. Checking back right now reveals that the average level of %CPU of the LLR workers has not improved, and notably the overall load average hasn't changed much. So, X11 is not the cause for the lower performance. Kernel on the 1P is 4.13.0-37-generic. The 2Ps run a 4.6 kernel (too lazy to update), except one which runs a 4.15 kernel without measurable performance difference.

Processors are the same, CPU clocks are the same (always the all-core turbo, or in case of PG LLR the all-core AVX turbo), number of memory channels per CPU (4) and memory clocks (DDR4-2400) are the same. But the 2Ps run registered memory, the 1P unregistered. Memory timings of the 1P should even be marginally better than on the 2Ps, but I'll have to check this eventually to be sure.

Gonna have to compare kernel configs some day.
Low-level userland libraries shouldn't make difference, since sllr64.3.8.20 is statically linked...

Edit,
sadly, I cannot properly quantify the performance difference yet, because none of my returned tasks was validated yet, naturally.
 
Last edited:

crashtech

Lifer
Jan 4, 2013
10,042
1,749
136
This Primegrid task just kills my F@H performance, maybe even worse that the last one where the two were run concurrently.
 

Ken g6

Programming Moderator, Elite Member
Moderator
Dec 11, 1999
15,664
2,762
55
Day 1 stats:

Rank___Credits____Username
1______1260204____xii5ku
45_____99548______crashtech
57_____73098______Howdy2u2
91_____39852______Ken_g6
115____33018______zzuupp
130____25746______Orange Kid

Rank__Credits____Team
2_____3068020____Czech National Team
3_____2830914____Aggie The Pew
4_____2391566____Sicituradastra.
5_____1531469____TeAm AnandTech
6_____802640_____Crunching@EVGA
7_____722053_____BOINC@Poland
8_____721750_____AMD Users



@StefanR5R is on fire! :) (Not literally, I hope!)
 

biodoc

Diamond Member
Dec 29, 2005
6,045
1,915
136
I put 6 cores from my 3930K on the race. Who knows maybe primegrid will be the choice for the FB sprint. :)
 

StefanR5R

Elite Member
Dec 10, 2016
4,628
5,572
136
Currently, the PrimeGrid web preferences say that ESP-LLR has 36 hours recent average CPU time. We are now 32 hours into the race. And even though run time ≠ CPU time, the next several hours may bring about some changes in team rankings and user rankings.
 
  • Like
Reactions: zzuupp

StefanR5R

Elite Member
Dec 10, 2016
4,628
5,572
136
So, just like in the previous race, my single-socket Xeon E5 (Linux Mint) is quite a bit slower than the dual-socket E5s (Gentoo and OpenSUSE). But this time the single-socket E5 isn't running any GPU jobs on the side.
It's not so bad as it seemed. The host has 8 validated tasks now, from which I calculate 130 kPPD if it was in steady state. The two 2P hosts which have the same processor are at 270 kPPD (2x 104 %) and 290 kPPD (2x 112 %, taking the 1P host as 100 % baseline). I.e. the lag of the Mint 1P behind the comparable Gentoo 2Ps is merely within the error of measurement.
 
  • Like
Reactions: TennesseeTony

bbhaag

Diamond Member
Jul 2, 2011
6,026
1,401
136
Does this race take longer to recognize stats than the last one we did? I started it at about 5:30 central time last night and I still don't see any of my contributions to the team.
 

Ken g6

Programming Moderator, Elite Member
Moderator
Dec 11, 1999
15,664
2,762
55
Does this race take longer to recognize stats than the last one we did? I started it at about 5:30 central time last night and I still don't see any of my contributions to the team.
I see you have eight tasks in progress, but no results returned yet. Did you configure your app_config.xml to handle ESP tasks like 321 tasks? You can just replace 321 with ESP in the file and select the menu option to read Config files.
 

bbhaag

Diamond Member
Jul 2, 2011
6,026
1,401
136
I see you have eight tasks in progress, but no results returned yet. Did you configure your app_config.xml to handle ESP tasks like 321 tasks? You can just replace 321 with ESP in the file and select the menu option to read Config files.
I don't have an app_config.xml to handle tasks. I didn't have one for the the 321 race. Do I need one for the ESP race? I guess I could create one. It looks pretty simple to do and straight forward.
 

zzuupp

Lifer
Jul 6, 2008
14,861
2,319
126
Currently, the PrimeGrid web preferences say that ESP-LLR has 36 hours recent average CPU time. We are now 32 hours into the race. And even though run time ≠ CPU time, the next several hours may bring about some changes in team rankings and user rankings.

36?? Single threaded, I'm running about twenty-ish.

Dual threaded first round was10-ish. Second round messed up timings. Third round thinks it will finish in under 10.
 

Ken g6

Programming Moderator, Elite Member
Moderator
Dec 11, 1999
15,664
2,762
55
I don't have an app_config.xml to handle tasks. I didn't have one for the the 321 race. Do I need one for the ESP race? I guess I could create one. It looks pretty simple to do and straight forward.
Not necessarily. I thought you had come in with one.

Anyway, you should probably either create an app_config.xml file or set your percent of cores used to 50%.
 

StefanR5R

Elite Member
Dec 10, 2016
4,628
5,572
136
@bbhaag, outside of the challenges I had tested llr321 and llr ESP on Haswell and on Broadwell-EP. Due to time constraints I didn't run the very same test regime for both applications, but what I saw confirmed the expectation that their performance profiles are similar (but not the same). For 4-core/8-thread Haswell Xeon E3, 3.4 GHz, 8 MB L3 cache, dual-channel DDR3,
  • llr321 throughput peaked when 1 task at a time was run with 8 threads (with almost no difference to 1 task with 7 threads), when no other processor load was present on the system,
  • llrESP throughput peaked at 1 task at a time with 6 threads (with almost no difference to 1 task with 7 threads), ditto without any other load present.
I estimate these optimum configs to give about 1.5 times the throughput compared with running 4 tasks with 1 thread each, and >1.5 times the throughput compared with running 8 tasks with 1 thread each, on this LGA1150 Haswell.

The next PrimeGrid challenges in June and August will have subprojects with much shorter runtimes per task. With those, multithreading will not scale as well as it does with llr321 and llrESP.

I made a thread with various PrimeGrid performance comparisons.

BTW, in my experience it is 100 % safe to change the multithreading configuration via app_config at any time during the execution of LLR tasks, regardless of completion percentage, and apply it to the still unfinished tasks. There are several ways to do that, but I'd recommend to (1.) shut down the boinc client, (2.) edit app_config.xml, (3.) restart the boinc client.

--------

@zzuupp, 20 hours sounds right for CPU time on a decent CPU with HT and MT off. But by 10 hours I believe you are referring to run time, not CPU time.

Before the challenge the web prefs showed even 41 hours recent average CPU time for llrESP. Apparently more modern machines were turned towards this project when the challenge started. However, CPU time does not only depend on the machine architecture, it also depends on the configuration of a given machine: Whether or not HT is used; whether the application was configured for multithreading, and to which thread count, and more factors.

As an example, the latest 20 tasks on one of my hosts with best throughput have averages of 4.4 hours run time but 43.2 hours CPU time. As you can see from that, both hyperthreading and multithreading are used, and an unusually high thread count at that.
 
  • Like
Reactions: bbhaag and zzuupp

Ken g6

Programming Moderator, Elite Member
Moderator
Dec 11, 1999
15,664
2,762
55
Day 2 stats:

Rank___Credits____Username
2______2603275____xii5ku
33_____291939_____crashtech
53_____193085_____Howdy2u2
94_____92746______zzuupp
97_____86340______Ken_g6
136____53162______bbhaag
155____45535______Orange Kid
221____6680_______biodoc

Rank__Credits____Team
2_____7198536____Czech National Team
3_____6932765____SETI.Germany
4_____5354821____Sicituradastra.
5_____3372765____TeAm AnandTech
6_____1866208____BOINC@Poland
7_____1639012____AMD Users
8_____1563316____Crunching@EVGA



Looks like the Formula BOINC sprint hasn't affected us too much here. From what I've seen of that project, people with Sandy Bridge or Ivy Bridge CPUs should really stick around here.
 

StefanR5R

Elite Member
Dec 10, 2016
4,628
5,572
136
I'd say between PrimeGrid/LLR and VGTU, people with Sandy Bridge or Ivy Bridge (or Bulldozer or Phenom, for that matter) are sitting between a rock and a hard place. :-(

And so do people with Haswell+: There is no easy choice which one to run. About the only ones who don't have to worry much about that are Ryzen owners...
 

Orange Kid

Elite Member
Oct 9, 1999
4,135
1,873
146
xii5ku has quite the battle going.:eek: He needs to throw everything he has at it and borg friends, enemies and neighbors. :):D
 
  • Like
Reactions: zzuupp

ASK THE COMMUNITY