1)
Message boards :
Number crunching :
Once in a Blue Moon Challenge
(Message 151238)
Posted 643 days ago by  roadrunner_gs
The answer to the posts of you both is: Your mileage may vary.
You could sometimes gain performance with SMT/HT enabled, you could sometimes lose performance.
It is not that deterministic and depends on the workload and CPU-architecture - so there is no "one answer fits all"-guideline.
Normally i would disable SMT/HT out of security reasons - but on my Thinkpad A285 with Ryzen 7 Pro 2700U for instance i found a bug with the nosmt-implementation where it draws more power and runs way hotter (75 °C vs 50 °C) with nosmt versus smt enabled, so i switched back.
|
2)
Message boards :
Number crunching :
Once in a Blue Moon Challenge
(Message 151200)
Posted 647 days ago by roadrunner_gs
If you go to go to the Challenge leaderboard and look for users that have 5 times or more
tasks than their neighbors, they're probably DC task hunting, and aborting a lot of tasks.
I don't think so.
As said, i have aborted around 10 % (now less) to reallign WUs (and that was a chore), so that thay are staggered an not end all around the same point and i am at 115 for around 1.4 million credits.
I am at position 59 and above me only one user (wscr) is sticking out with around 350 WUs @ 1.7 million credits compared to the surroundings and according to some of his hosts is or was actively doing this.
|
3)
Message boards :
Number crunching :
Once in a Blue Moon Challenge
(Message 151192)
Posted 648 days ago by roadrunner_gs
I only took the disk and put it in the other rig.
Done this time and again with different discs, boinc never created new hosts.
So i presume i am fine.
|
4)
Message boards :
Number crunching :
Once in a Blue Moon Challenge
(Message 151188)
Posted 648 days ago by roadrunner_gs
(...)
This computer probably isn't even set up to run the challenge as it's doing this on many of the sub-projects. It's doing more PSP than others because the server is set to favor challenge tasks (this improves performance.)
So there you have it. The big number of aborts is mostly due to a single broken computer.
Interesting - and good explanation which with lacking insight i could not deduce.
by the way: I changed my E5-2690 out and my E5-2670 v3 in, now 2 tasks were started with AVX and are completed with FMA3, is this a game breaker? Should i abort the tasks, or would they come out just fine?
(since nobody asked - i swapped the ssd, since the CPUs won't fit in the same board. ;) )
|
5)
Message boards :
Number crunching :
Once in a Blue Moon Challenge
(Message 151178)
Posted 648 days ago by roadrunner_gs
Why is the abortion-rate that high?
I mean, i too aborted around 10 %, to "realign" the WUs with some DC-tasks, but that high a percentage?
Any ideas?
|
6)
Message boards :
Number crunching :
Once in a Blue Moon Challenge
(Message 151163)
Posted 650 days ago by roadrunner_gs
Core temperatur is just fine and the frequency is only 100 MHz below the lowest max frequency.
I could not see the Package Power or Core Power and would presume those are the restricting factors.
My E5-2690 also clocks down from its 3.3 GHz TurboBoost to around 3.1/3.2 since it draws 145 W with the former and down to the specced 135 W with the later. (~85 °C, borderline ok ;) water cooling would help i think, but i would not sink more money into this rig).
On your front it could be a faulty memory module or (a rare) faulty CPU(-part).
|
7)
Message boards :
Number crunching :
Once in a Blue Moon Challenge
(Message 151149)
Posted 650 days ago by roadrunner_gs
Your CPU is running just fine and even with SMT/HT around 100% faster (?) than my Ryzen 7 3700X wit SMT/HT deactivated - so you should be fine. Albeit you are running Linux and me Windows 10, i am running 2 x 4 Threads and you are running 2 x 8 Threads.
The WU-scheduling "should" (with a large grain of salt) be handled via the process scheduler in Linux - it is HT/CCX-aware to a better degree than you could be with manually pinning WUs to cores.
Core-topology is normaly stored somewhere around /sys/devices/system/cpu/cpu${number}/topology/core_siblings_list
numactl --hardware should give you an insight if the kernel thinks some "cores" (more like domains) have fancy relationship towards each other (but your lscpu shows one numa-node, so that is out of the question).
The MT-settings are a bit coarse:
I got 1 4-Core, 2 8-Core, 1 2x8-Core and 1 12-Core (not running at the moment), so 4 Threads is a safe bet for me.
I clocked several Xeon (E5 v1 to v3, E7, Silver/Bronze/Gold) with 321-llr and i think 3k-FFTs but i couldn't fully get a clear speedup recomendation regarding cache-sizes.
For disabling SMT/HT i would go the EFI/BIOS-route, or adding nosmt as a grub2 kernelopts-parameter.
|
8)
Message boards :
Number crunching :
World Emoji Day Challenge
(Message 150871)
Posted 677 days ago by roadrunner_gs
Somehow i found my RX 580 crunching on this WU for 1 day and 10 hours, i aborted it.
Next picked WU exhibits normal runtime.
Curious.
|
9)
Message boards :
Number crunching :
World Emoji Day Challenge
(Message 150858)
Posted 678 days ago by roadrunner_gs
Nice 3090 there Nick!
I was surprised to see the lack of 1st it's was getting but I'm sure there's a good reason for that. I got 84% 1st with a 1660 super, so a 3090 should be abve 90%?
LOL... a 1660 is faster than my Radeon VII :D
Same here GTX 1650 faster then AMD RX 580. o_O
I knew this from AP27, but i would have thought the AMD was faster here.
|
10)
Message boards :
Number crunching :
World Emoji Day Challenge
(Message 150831)
Posted 679 days ago by roadrunner_gs
It Appears that all of my openclatiGFN17LOW tasks are being reported as cpuGFN17LOW tasks. Anyone know why?
Are they really running as/at CPU?
See my post two beĺow: AMD Ryzen 3700x with AMD RX 580, only getting openclATI-Tasks and reporting as such?
Okay, my host is getting work now for GFN17low@CPU
|
