New PC not performing as well as expected

Message boards : Number crunching : New PC not performing as well as expected
Message board moderation

To post messages, you must log in.

Previous · 1 · 2 · 3 · Next

AuthorMessage
Ian&Steve C.
Avatar

Send message
Joined: 28 Sep 99
Posts: 4267
Credit: 1,282,604,591
RAC: 6,640
United States
Message 2016802 - Posted: 27 Oct 2019, 1:25:20 UTC - in response to Message 2016800.  

Other things competing for CPU resources show up as a discrepancy between CPU time and Runtime.


and i see that.

much larger discrepancy between cpu and run times than someone else running properly.
Seti@Home classic workunits: 29,492 CPU time: 134,419 hours

ID: 2016802 · Report as offensive
Profile Keith Myers Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 13164
Credit: 1,160,866,277
RAC: 1,873
United States
Message 2016803 - Posted: 27 Oct 2019, 1:26:24 UTC - in response to Message 2016800.  

However- how many modules make up that amount of RAM? Are they in the correct memory slots?

Good point. We had a new user and new to hardware that didn't realize that a quad channel system does not run correctly when the memory is put into the wrong slots recently.
He also had atrocious run times until the memory configuration was fixed.
Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)
ID: 2016803 · Report as offensive
Grant (SSSF)
Volunteer tester

Send message
Joined: 19 Aug 99
Posts: 13958
Credit: 208,696,464
RAC: 304
Australia
Message 2016804 - Posted: 27 Oct 2019, 1:26:55 UTC - in response to Message 2016801.  
Last modified: 27 Oct 2019, 1:27:39 UTC

Wasn't there a LINUX distro that had issues with power saving kicking in when it shouldn't have some time back? Could this be something like that?
Shouldn't be an issue. He is running a very recent kernel with all the latest cpu scheduler improvements.
Yeah, but all I can think of is low CPU clock speed. Even memory modules in the wrong slots couldn't cause such a huge performance hit for a single socket dual memory channel system, could it?
It would explain the poor CPU performance, and the less than expected GPU performance for the given hardware & applications.
Grant
Darwin NT
ID: 2016804 · Report as offensive
Ian&Steve C.
Avatar

Send message
Joined: 28 Sep 99
Posts: 4267
Credit: 1,282,604,591
RAC: 6,640
United States
Message 2016806 - Posted: 27 Oct 2019, 1:30:38 UTC - in response to Message 2016804.  

the clock speed is measured and reported. so it's not that.
Seti@Home classic workunits: 29,492 CPU time: 134,419 hours

ID: 2016806 · Report as offensive
Grant (SSSF)
Volunteer tester

Send message
Joined: 19 Aug 99
Posts: 13958
Credit: 208,696,464
RAC: 304
Australia
Message 2016807 - Posted: 27 Oct 2019, 1:43:33 UTC - in response to Message 2016806.  

the clock speed is measured and reported. so it's not that.
Reported at the time the stderr.txt log was produced. If it is the actual clock speed, not just the specs for that CPU, it would still be for at the time the stderr.txt log was produced, not necessarily for the period the WU was being processed.
It would be nice to have a realtime monitor reporting the CPU frequency to be sure.

And if it is running at that speed, then why the hell are things taking so long to process? (as I pointed out it's not due to other processes taking up CPU time, although there is some of that going on). And i'd be very surprised if a dual channel memory system would have such poor performance from the wrong memory slots being used.
Grant
Darwin NT
ID: 2016807 · Report as offensive
Profile Keith Myers Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 13164
Credit: 1,160,866,277
RAC: 1,873
United States
Message 2016808 - Posted: 27 Oct 2019, 1:48:37 UTC - in response to Message 2016807.  

Reported at the time the stderr.txt log was produced.

That's an important point. I often see very low cpu speed reported in some of my cpu tasks. Yet I know that all cores run continuously at a fixed cpu clock set in the BIOS. So the reason for the low stderr.txt clock speed output can only be reasoned as it was sampled just as the load left the core and before the next task load was picked up. I have C-states enabled in the BIOS. The OS is opportunistic in dropping a cpu core clock when it detects that it has very little compute load on it. But the clocks ramp up to the set BIOS speed when the compute load persists.
Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)
ID: 2016808 · Report as offensive
Profile Keith Myers Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 13164
Credit: 1,160,866,277
RAC: 1,873
United States
Message 2016810 - Posted: 27 Oct 2019, 1:54:14 UTC - in response to Message 2016807.  

And i'd be very surprised if a dual channel memory system would have such poor performance from the wrong memory slots being used.

Ha ha LOL. Over in the OCN overclocking threads, the first question we ask when someone reports their memory only runs at 2133Mhz and they are unable to even run at XMP spec speed for the modules, is "Are the memory modules installed into the recommended slots per the motherboard manual?" It does make a difference owing to whatever memory topology the motherboard is designed to. There are a lot of people that are very inexperienced builders. And worse, there is a lot of shoddy work done by system builders who should know better.
Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)
ID: 2016810 · Report as offensive
Ian&Steve C.
Avatar

Send message
Joined: 28 Sep 99
Posts: 4267
Credit: 1,282,604,591
RAC: 6,640
United States
Message 2016812 - Posted: 27 Oct 2019, 3:07:45 UTC - in response to Message 2016807.  

Why would it clock FASTER when writing a file vs processing a work unit? Think about that. That makes no sense. I can see the opposite like Keith says, since the core ramps down for less intensive loads. But not the other way around.

I’m confident that clock speeds are not the issue. It’s something else.
Seti@Home classic workunits: 29,492 CPU time: 134,419 hours

ID: 2016812 · Report as offensive
Grant (SSSF)
Volunteer tester

Send message
Joined: 19 Aug 99
Posts: 13958
Credit: 208,696,464
RAC: 304
Australia
Message 2016820 - Posted: 27 Oct 2019, 4:02:12 UTC - in response to Message 2016812.  
Last modified: 27 Oct 2019, 4:06:48 UTC

Why would it clock FASTER when writing a file vs processing a work unit? Think about that. That makes no sense. I can see the opposite like Keith says, since the core ramps down for less intensive loads. But not the other way around.
I can see it happening as I've suggested, as it happens on notebooks & the like all the time with heavy CPU loads- Thermal throttling when under extended loads. That's if there is a thermal issue. If the data in the stderr_txt file is the reported sensor output (not the CPU's specified abilities) clock speed at the time the CPU is queried for the data to go in the stderr_txt file, but that is just a brief period- not the whole time the WU is being processed.
Maybe someone responsible for application development would care to fill us in on just where the stderr_txt CPU information comes from- current CPU sensor data, or just the CPUs specs (from Keith's post it would seem to be reported sensor data)- and when it's gathered?
Grant
Darwin NT
ID: 2016820 · Report as offensive
Profile Keith Myers Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 13164
Credit: 1,160,866,277
RAC: 1,873
United States
Message 2016832 - Posted: 27 Oct 2019, 5:47:27 UTC - in response to Message 2016820.  
Last modified: 27 Oct 2019, 5:49:53 UTC

I have a simple script file that I can run to watch the cpu core clocks. It normally polls every second. During crunching on all the cores, the clocks are locked in solid down to the hundredths of a Mhz to the all-core multiplier speed I set in the BIOS. But I also do not run all cores in use and I do not lock the cpu application to any one specific core so the load on any one core can move around. I can actually lock the load on an Intel cpu but because of the unique method that AMD uses to migrate loads around based on its internal microarchitecture, trying to lock loads on a Ryzen or later cpu works against the cpu clocks and load balancing. So I can easily watch loads move off one core graphically in my GKrellm monitor cpu panels. If the load drops off a core then the C-state transitions in a matter of microseconds down to C6 state and the core clock briefly dips. If the task is finishing up and writing out the stderr.txt file, I can see how the snapshot of the core clock at the instant in time gets captured at the lower clock speed. Does not mean the whole task ran at that clock speed for the entirety of the duration to crunch the task.

Every 1.0s: cat /proc/cpuinfo | grep "^[c... Serenity: Sat Oct 26 22:43:02 2019

cpu MHz : 4148.875
cpu MHz : 4150.141
cpu MHz : 4149.151
cpu MHz : 4150.137
cpu MHz : 4150.138
cpu MHz : 4150.139
cpu MHz : 4150.140
cpu MHz : 4149.780
cpu MHz : 4150.104
cpu MHz : 4150.139
cpu MHz : 4150.139
cpu MHz : 4150.063
cpu MHz : 4150.140
cpu MHz : 4150.140
cpu MHz : 4149.933
cpu MHz : 4150.109
cpu MHz : 4150.138
cpu MHz : 4150.065
cpu MHz : 4149.617
cpu MHz : 4150.137
cpu MHz : 4150.138
cpu MHz : 4150.138
cpu MHz : 4150.139
cpu MHz : 4150.139

I do not have any thermal throttling as I have custom water cooling. Enough that there is plenty of thermal headroom to prevent the cpu from ever thermal throttling the cpu clocks to reduce power and heat.
Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)
ID: 2016832 · Report as offensive
Profile Keith Myers Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 13164
Credit: 1,160,866,277
RAC: 1,873
United States
Message 2016835 - Posted: 27 Oct 2019, 6:06:05 UTC

Maybe someone responsible for application development would care to fill us in on just where the stderr_txt CPU information comes from- current CPU sensor data, or just the CPUs specs (from Keith's post it would seem to be reported sensor data)- and when it's gathered?


Pretty sure it is coming from here:

https://github.com/BOINC/boinc/blob/8a14c59fb12e053474276bb8edb39798f4417760/client/hostinfo_unix.cpp#L446

Since the normal way to probe the core clocks is to look at /proc/cpuinfo as in my previous post, the client probably looks there too for the cpu core clock to print out the header information of the stderr.txt.

You might have noticed that only the Linux apps print out the cpu core speed in the stderr.txt. The Windows apps do not.
Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)
ID: 2016835 · Report as offensive
Grant (SSSF)
Volunteer tester

Send message
Joined: 19 Aug 99
Posts: 13958
Credit: 208,696,464
RAC: 304
Australia
Message 2016837 - Posted: 27 Oct 2019, 6:48:13 UTC - in response to Message 2016835.  

You might have noticed that only the Linux apps print out the cpu core speed in the stderr.txt. The Windows apps do not.
Yep.


What we know & don't know.

Extremely long CPU processing times, for a CPU of it's series, using the current AVX application.
Up to date OS.
Not thermal throttling (had some PMs with Eric B- CPU is high 50°c to just over 60°c; no idea of ambient temp)- Low CPU temperatures indicative of the CPU not working hard.
Reported clock sped at completion of WU inline with CPU capabilities, but what about during the processing of the WU?
Poor computation times not due to other programme contention for CPU time or use of "Use at most xxx % of CPU time" setting as that results in a very large difference between CPU time and Run time and that isn't the case here (yes there is more than the usual difference between them- but the Run time is excessive, which does not occur with programme contention or "Use at most xxx % of CPU time" limits set).
Memory module placement unknown- but would wrong slots being used really have such a significant impact on a single socket dual channel system? (multi socket quad channels- certainly, but single socket dual channel?).
Grant
Darwin NT
ID: 2016837 · Report as offensive
Ian&Steve C.
Avatar

Send message
Joined: 28 Sep 99
Posts: 4267
Credit: 1,282,604,591
RAC: 6,640
United States
Message 2016849 - Posted: 27 Oct 2019, 12:24:29 UTC

Just have him install and run the i7z package in a terminal while the CPU jobs are running.

That will definitely say what clock speeds it’s running. It will list a per-thread clock speed.
Seti@Home classic workunits: 29,492 CPU time: 134,419 hours

ID: 2016849 · Report as offensive
Profile Keith Myers Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 13164
Credit: 1,160,866,277
RAC: 1,873
United States
Message 2016860 - Posted: 27 Oct 2019, 15:05:09 UTC - in response to Message 2016849.  
Last modified: 27 Oct 2019, 15:05:25 UTC

Just have him install and run the i7z package in a terminal while the CPU jobs are running.

That will definitely say what clock speeds it’s running. It will list a per-thread clock speed.

Yes for Intel cpus that app is the best. Also shows you the Vcore on each core and whether the core is moving in and out of C-states while crunching. If the core crunching a task doesn't stay 100% of the time in C0 state, then that is a problem. If the cores are moving from C0 to C1 state, then something is causing Halts on that core due to power saving.
Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)
ID: 2016860 · Report as offensive
Grant (SSSF)
Volunteer tester

Send message
Joined: 19 Aug 99
Posts: 13958
Credit: 208,696,464
RAC: 304
Australia
Message 2016950 - Posted: 28 Oct 2019, 7:06:42 UTC
Last modified: 28 Oct 2019, 7:17:14 UTC

Ian&Steve C called it.
CPU Clock speeds Ok.

Guess what?
Memory module in the wrong slot on a dual channel system can have a huge performance impact.
Even so, performance is still down on what i'd expect from that CPU- it should be outperforming the other system by a healthy margin.
i9-9900K APR 34.79 GFLOPS
i7-3960X APR 36.92 GFLOPS

Need to check memory speed & timing?

Or better yet, wait a while and let things settle down. The newer system, although it has a lower APR at present, appears to be processing similar WUs approx 15% faster than the older system.
Grant
Darwin NT
ID: 2016950 · Report as offensive
Ian&Steve C.
Avatar

Send message
Joined: 28 Sep 99
Posts: 4267
Credit: 1,282,604,591
RAC: 6,640
United States
Message 2016988 - Posted: 28 Oct 2019, 15:04:28 UTC - in response to Message 2016950.  

what motherboard is he running?
what memory kit is he running? (speeds and timings, exact part number would help)
is all the memory matching? or is it a mix of brands/speeds?

it's possible he just needs to enable the XMP profile in the BIOS. but Intel chips usually aren't performance limited by the RAM like AMD chips are.
Seti@Home classic workunits: 29,492 CPU time: 134,419 hours

ID: 2016988 · Report as offensive
Profile Keith Myers Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 13164
Credit: 1,160,866,277
RAC: 1,873
United States
Message 2017000 - Posted: 28 Oct 2019, 15:58:29 UTC

it's possible he just needs to enable the XMP profile in the BIOS. but Intel chips usually aren't performance limited by the RAM like AMD chips are.

Not as much as AMD but there still is a degradation. I can only run 2400Mhz on my 3000 XMP memory with my engineering sample Xeon E5-1660 V4(i7-6850K equivalent other than an extra two cores) running at 4Ghz. I can run the memory at XMP 3000 spec with the i7-6850K running at the same 4Ghz and the tasks are much faster just because of the memory speed improvement.
Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)
ID: 2017000 · Report as offensive
Ian&Steve C.
Avatar

Send message
Joined: 28 Sep 99
Posts: 4267
Credit: 1,282,604,591
RAC: 6,640
United States
Message 2017007 - Posted: 28 Oct 2019, 16:43:53 UTC - in response to Message 2017000.  

I think the biggest difference in those cases where you’re not running XMP is the timings rather than the actual clock speed. XMP will run the spec timings. But without it (and without manual adjustment) it defaults to slow speeds and very loose timings.
Seti@Home classic workunits: 29,492 CPU time: 134,419 hours

ID: 2017007 · Report as offensive
Profile Keith Myers Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 13164
Credit: 1,160,866,277
RAC: 1,873
United States
Message 2017013 - Posted: 28 Oct 2019, 17:53:42 UTC - in response to Message 2017007.  

In my case no. The 2400 settings were the same primary and secondary timings as XMP at 3000. You know I pay attention to those things to get the best performance out of my rigs.
Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)
ID: 2017013 · Report as offensive
Ian&Steve C.
Avatar

Send message
Joined: 28 Sep 99
Posts: 4267
Credit: 1,282,604,591
RAC: 6,640
United States
Message 2017014 - Posted: 28 Oct 2019, 18:03:09 UTC - in response to Message 2017013.  

Maybe in your case it comes down to using an ES chip. They sometimes have weird issues.

I’ve never seen a huge difference with slower mem clock speeds.

One of my older systems (no longer in service) was running DDR3-1600 mem with the CPU at like 3GHz and it still churned out WUs in under an hour. Memory speeds usually don’t matter too much on Intel chips, usually only in things like memory benchmarks do you see the differences that scale with clock speeds.
Seti@Home classic workunits: 29,492 CPU time: 134,419 hours

ID: 2017014 · Report as offensive
Previous · 1 · 2 · 3 · Next

Message boards : Number crunching : New PC not performing as well as expected


 
©2025 University of California
 
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.