[Xenomai] Latency test results Xeon E5-1650 higher than Xeon E5620

[Xenomai] Latency test results Xeon E5-1650 higher than Xeon E5620

[Beaufils Sylvie (IT&LAB)]

Hi all,

We run linux kernel 3.2.34 + ipipe-core-3.2.21-x86-1 + xenomai 2.6.1 on two platforms, one with Xeon E5620 and the other with Xeon E5-1650.

E5-1650 shows higher latencies than E5620 when running the following test on an idle system:

latency -t 2 -T60 -q ; latency -t 1 -T60 -q ; latency -t 0 -T60 -q

++++++ Xeon E5-1650 ++++++
== Test mode: in-kernel timer handler
---|-----------|-----------|-----------|--------|------|-------------------------
RTS|     -3.156|     -2.073|     10.554|       0|     0|    00:01:00/00:01:00
== Test mode: in-kernel periodic task
---|-----------|-----------|-----------|--------|------|-------------------------
RTS|     -2.694|     -1.215|     16.683|       0|     0|    00:01:00/00:01:00
== Test mode: periodic user-mode task
---|-----------|-----------|-----------|--------|------|-------------------------
RTS|     -2.445|     -0.412|     15.308|       0|     0|    00:01:00/00:01:00

++++++ Xeon E5620 ++++++
== Test mode: in-kernel timer handler
---|-----------|-----------|-----------|--------|------|-------------------------
RTS|     -3.759|     -2.669|      1.056|       0|     0|    00:01:00/00:01:00
== Test mode: in-kernel periodic task
---|-----------|-----------|-----------|--------|------|-------------------------
RTS|     -2.656|     -1.567|      6.852|       0|     0|    00:01:00/00:01:00
== Test mode: periodic user-mode task
---|-----------|-----------|-----------|--------|------|-------------------------
RTS|     -2.060|     -0.964|      7.631|       0|     0|    00:01:00/00:01:00


I have read that "The most common reason for high latencies on x86 are SMIs". But the SMI workaround does not manage the Patsburg LPC controller. I tried to add the PCI ID [8086:1d41] to ksrc/arch/x86/smi.c with this result:
[    1.926417] Xenomai: SMI-enabled chipset found
[    1.926490] Xenomai: SMI workaround failed!

Should the SMI workaround be managed for the C602 chipset ? Could it explain the above latency results ?

Thanks,
Sylvie

Re: [Xenomai] Latency test results Xeon E5-1650 higher than Xeon E5620

[Gilles Chanteperdrix]

On 12/11/2012 06:49 PM, Beaufils Sylvie (IT&LAB) wrote:
> I have read that "The most common reason for high latencies on x86

> are SMIs". But the SMI workaround does not manage the Patsburg LPC
> controller. I tried to add the PCI ID [8086:1d41] to
> ksrc/arch/x86/smi.c with this result: [    1.926417] Xenomai:
> SMI-enabled chipset found [    1.926490] Xenomai: SMI workaround
> failed!
> 
> Should the SMI workaround be managed for the C602 chipset ? Could it
> explain the above latency results ?


Maybe, but as mentioned in the part of the troubleshooting guide you do
not quote, a "pathological" latency is a latency around say 100us.

The result of a latency test running:
- 1 minute
- on an idle system

is absolutely meaningless.

Anyway, if you want to understand the difference, you may use the I-pipe
tracer.

-- 
                                                                Gilles.