CPU utilization and hyperthreading

CPU utilization and hyperthreading

[erezz]

Hi,

I'm using a 2.6.4-52 (Suse 9.1 pro distribution) kernel on a 3.2 dual
XEON machine (with hyperthreading).

I'm trying to analyze the performance of a kernel module that contains
4 kernel threads (3 of them are more active). Also, some of the work is
done in a tasklet. The CPU load (as shown in vmstat) doesn't reach more
than 50%.  When trying to artificially stress the system with some
useless loops, the CPU load reaches 65% (but I cannot reach 100%). Also, the
interrupts count in vmstat is around 20000 (1000 when idle).
 
My questions are:
1. Why can't I reach 100%? 
2. I guess that since I have 4 virtual CPUs (as shown in /proc/cpuinfo), not all
of them are stressed. Is there a way to see the CPU utilization for each CPU
(vmstat doesn't allow that)? 
3. When I load a module with a kernel thread that runs an infinite loop, I get
25% CPU utilization (50% if I run 2 threads). Is it because I'm using 100% of a
single CPU (out of 4)?
4. How is the time spent in interrupt context estimated? Is it possible to view
it somehow?

Thanks
Erez

Re: CPU utilization and hyperthreading

[Denis Vlasenko]

On Sunday 06 November 2005 16:13, erezz@techunix.technion.ac.il wrote:
> Hi,
> 
> I'm using a 2.6.4-52 (Suse 9.1 pro distribution) kernel on a 3.2 dual
> XEON machine (with hyperthreading).
> 
> I'm trying to analyze the performance of a kernel module that contains
> 4 kernel threads (3 of them are more active). Also, some of the work is
> done in a tasklet. The CPU load (as shown in vmstat) doesn't reach more
> than 50%.  When trying to artificially stress the system with some
> useless loops, the CPU load reaches 65% (but I cannot reach 100%). Also, the
> interrupts count in vmstat is around 20000 (1000 when idle).
>  
> My questions are:
> 1. Why can't I reach 100%? 

You can.

# while true; do true; done &
# while true; do true; done &
# while true; do true; done &
# while true; do true; done &

> 2. I guess that since I have 4 virtual CPUs (as shown in /proc/cpuinfo), not all
> of them are stressed. Is there a way to see the CPU utilization for each CPU
> (vmstat doesn't allow that)? 

top

> 3. When I load a module with a kernel thread that runs an infinite loop, I get
> 25% CPU utilization (50% if I run 2 threads). Is it because I'm using 100% of a
> single CPU (out of 4)?
> 4. How is the time spent in interrupt context estimated? Is it possible to view
> it somehow?
--
vda