System with large fluctuations in CPU load and SMP client

A

APOLLO

[H]ard|DCer of the Month - March 2009
Joined
Sep 17, 2000
Messages
9,089
I have a system that is exhibiting a bizarre CPU load profile with the SMP client. I tried Windows and Linux SMP clients and they both demonstrated the same wild fluctuations in CPU load. Basically, the graph in Task Manager shows unsteady, jagged lines and Core Temp application does not report a continuous 100% load on any core. The percentages jumps like crazy between 100% and much lower loads percentages.

Furthermore, temps are very low on each core of both processors. Performance in the SMP client is nearly half what it should be compared to other systems with similar specs in my farm. It's definitely not memory, I tried so many modules it cannot be anything related to that. Anyone see this or can recommend what I can do to test the system for problems?
 
Last edited:
post your config and see if maybe you accidently change the cpu usage.. even at 99% cpu usage it would cause weird fluctuations in cpu usage.. could also be that the cpu's unstable.. needing more voltage..
 
Two ideas, both related to power:

1. Make sure Windows power management is set to "Performance."

2. Try a different PSU.

I'm sure that you've been through the bios and turned off all the power management stuff, but that's another thing that comes to mind.
 
Very odd Apollo and even stranger that you are experiencing the same problem under both Windows and Linux. *scratches head*
 
Mr. Pedantic said:
And check your CPU temps as well.
Click to expand...
Temps are very low.

Tobit said:
Very odd Apollo and even stranger that you are experiencing the same problem under both Windows and Linux. *scratches head*
Click to expand...
It's definitely hardware but not sure what. Almost like thermal management has kicked in somehow but all the settings are disabled in the BIOS.
 
djstarfox said:
What is the model and speed of your CPU(s)?
Click to expand...
Dual ES X5340 Xeons

What is the CPU graph result using this tiny little program called cpuburn-in?
Click to expand...
That program looks like it is for older processors. I have 8 cores, how am I to properly use this? Also, there are warnings about possible damage running it.
 
APOLLO said:
That program looks like it is for older processors. I have 8 cores, how am I to properly use this? Also, there are warnings about possible damage running it.
Click to expand...
Hmm... sorry about the link. I guess they removed the x86 architecture code. Anyway, some kind of CPU benchmark should not exhibit bursty CPU usage. There's a command line program called stress (stress -t 30 -c 8) for Linux and hundreds of CPU benchmarks for Windows.

While the benchmark or SMP client are running, do a quick Linux command to make sure they're running at full speed:
grep MHz /proc/cpuinfo

You could make this real quick by locking the SMP client's CPU affinity with 'taskset'. You can also do this in Windows via the Task Manager (in the processes list).

When you originally said the CPU is bursty, I thought that was because your SMP client program is not using all cores. Windows kernel will alternate cores quickly in an attempt to space out the CPU load; this will show up as lots of spikes for each CPU. I suspect your CPUs are fine, and it's just this SMP client that's using your CPU in a weird way.

Hope that helps.

PS Had to correct 'busty' to 'bursty' a few times. Freudian slip I suppose. ;)
 
djstarfox said:
Hmm... sorry about the link. I guess they removed the x86 architecture code. Anyway, some kind of CPU benchmark should not exhibit bursty CPU usage. There's a command line program called stress (stress -t 30 -c 8) for Linux and hundreds of CPU benchmarks for Windows.
Click to expand...
I guess I could look into some benchmarking tools for Windows. I just don't know which would be best to determine the underlying cause for this issue. I rarely if ever need to benchmark anything in my systems.

While the benchmark or SMP client are running, do a quick Linux command to make sure they're running at full speed:
grep MHz /proc/cpuinfo
Click to expand...
This command listed the current speed of each core and it was indeed the full speed I had set it at.

You could make this real quick by locking the SMP client's CPU affinity with 'taskset'. You can also do this in Windows via the Task Manager (in the processes list).
Click to expand...
I'm using Bill's Process Manager. The cores are fixed to the SMP client this way.

When you originally said the CPU is bursty, I thought that was because your SMP client program is not using all cores. Windows kernel will alternate cores quickly in an attempt to space out the CPU load; this will show up as lots of spikes for each CPU. I suspect your CPUs are fine, and it's just this SMP client that's using your CPU in a weird way.
Click to expand...
I have several systems with similar specifications. This system is definitely underperforming. We're talking at least 30-40% below the norm for this system's performance levels. Another thing I didn't mention in the OP is reducing the number of cores allocated for the SMP client from the maximum of 8 to around 6 cores yields superior performance. Meaning, I can't even utilize all the cores available without negatively impacting performance so something definitely is wrong.
 
sounds like something isnt stable then.. either the power source, or cpu voltage..

good program to try is cinbench 11.5 it will run long enough that you can see the cpu usage and it will use all 8 cores the entire time.. then you will know for sure what the problem is..
 
as monkey says, try cinebench.....probably more volts to the CPU are needed
 
I'm going to give Cinebench a try later today, but this is a server board - no voltage tweaks allowed unfortunately. :(
 
well then what are you waiting for? get a(nother) SR-2!!!!!!!!!!!
 
Now that I have a better idea what's happening, I agree with the first replies that say this is a power supply issue. Low temps under load give even more evidence of this. Surprised you don't get random lockups.
 
djstarfox said:
Now that I have a better idea what's happening, I agree with the first replies that say this is a power supply issue. Low temps under load give even more evidence of this. Surprised you don't get random lockups.
Click to expand...
I have a GX2 recently installed and it can very well be causing power fluctuations. Hmm, I'm going to see what I can do if that is this case. I don't want to get a new PSU for this system. It's too old to upgrade.
 
Normally, a server-class board will run longer without problems than a desktop board. However, the MOSFET on motherboards does wear out, causing voltage droop. You should be able to monitor this with voltage monitor software such as Speedfan.
 
Sorry for taking long to update this thread. I ran the CPU test in Cinebench and I'm not certain what I was supposed to note. It received a 5.85pts score on the CPU test and the results were subpar for the specs. This was expected since I'm not seeing a steady 100% load on the processors and that will undoubtedly prevent turning in a proper performance run.
 
You have the memory matched up in regards to ecc and registered?
You have it properly allocated per CPU and per channel per cpu if it is that kind of setup?
I would strip the ram down to min to boot and see what you get on cinebench.
Socket 771 xeons
http://ark.intel.com/Product.aspx?id=33081
if so it calls out a 67 degree limit? Do you have all the baffles this system should have?

The other thought if you haven't already checked is to make sure the heatsinks are on properly as it could be running throttled due to the temp spikes to high whenever they come out of throttling.

Lastly try running just one cpu (or 2 of 4 if four socket) and see what you get. So this is like a HP DL 380 or ML370 of the G5 series? Or something equivalent?

Did all the ram get detected in the Operating systems you were using?
 
nwtarget,

Answer is affirmative to everything. I have checked all that I could without removing processors or cards yet (no time). The system is in an open case with massive solid copper sinks and huge alpha fans that sound like jet engines. I had also tested a plethora of RAM combinations and different modules so it isn't memory. This is a custom built system and everything checks fine except for CPU usage. These are 'extended specs' processors and don't know if that might be related to the problem. Best guess so far is the PSU being insufficient to the task after I installed the GX2. When I had a single 8800GT there wasn't an issue. However, shutting off the GPU clients and just leaving the SMP client running does not improve anything. :confused:
 
APOLLO said:
nwtarget,

Answer is affirmative to everything. I have checked all that I could without removing processors or cards yet (no time). The system is in an open case with massive solid copper sinks and huge alpha fans that sound like jet engines. I had also tested a plethora of RAM combinations and different modules so it isn't memory. This is a custom built system and everything checks fine except for CPU usage. These are 'extended specs' processors and don't know if that might be related to the problem. Best guess so far is the PSU being insufficient to the task after I installed the GX2. When I had a single 8800GT there wasn't an issue. However, shutting off the GPU clients and just leaving the SMP client running does not improve anything. :confused:
Click to expand...


put the 8800GT back in and see if it solves the problem.. the 9800GX2 probably uses the same if not more power idle then a 8800GT does full load..

only other thing it could be is that your overloading one of the rails but no clue without knowing the psu and if your using a bunch of molex to 8pin/6pin adapters..
 
you might be heading in to two PSU country!

As you no doubt know HK runs up to THREE on some of his rigs.
 
nwrtarget said:
you might be heading in to two PSU country!

As you no doubt know HK runs up to THREE on some of his rigs.
Click to expand...
I have run dual PSUs for GPUs in the past. There is a spare I'm going to try running for the card only. If that doesn't work then I'll remove the suspect card like sirmonkey suggested.
 
You must log in or register to reply here.
Back
Top