Difficulties with cyclictest on AMD CPU

Difficulties with cyclictest on AMD CPU

[Stefan Jenisch]

Hello folks!

I am having some difficulties with cyclictest on an “AMD GX-210HA SOC 
with Radeon(tm) HD Graphics” CPU. Latency values are better with some 
CPU load (execution of an endless loop) than without.

When running the command “cyclictest -c 0 -t15 -p 95 -i 500 -n -l 
100000” I get the output:

policy: fifo: loadavg: 0.25 0.09 0.06 1/201 2349

T: 0 ( 2335) P:95 I:500 C: 100000 Min: 11 Act: 19 Avg: 21 Max: 231

T: 1 ( 2336) P:95 I:1000 C: 50000 Min: 10 Act: 13 Avg: 23 Max: 222

T: 2 ( 2337) P:95 I:1500 C: 33333 Min: 11 Act: 19 Avg: 24 Max: 198

T: 3 ( 2338) P:95 I:2000 C: 25000 Min: 10 Act: 16 Avg: 25 Max: 185

T: 4 ( 2339) P:95 I:2500 C: 20000 Min: 12 Act: 22 Avg: 25 Max: 196

T: 5 ( 2340) P:95 I:3000 C: 16666 Min: 11 Act: 20 Avg: 23 Max: 233

T: 6 ( 2341) P:95 I:3500 C: 14285 Min: 12 Act: 19 Avg: 25 Max: 209

T: 7 ( 2342) P:95 I:4000 C: 12500 Min: 10 Act: 13 Avg: 25 Max: 123

T: 8 ( 2343) P:95 I:4500 C: 11111 Min: 11 Act: 65 Avg: 29 Max: 203

T: 9 ( 2344) P:95 I:5000 C: 10000 Min: 11 Act: 26 Avg: 22 Max: 175

T:10 ( 2345) P:95 I:5500 C: 9091 Min: 13 Act: 30 Avg: 31 Max: 165

T:11 ( 2346) P:95 I:6000 C: 8333 Min: 12 Act: 19 Avg: 23 Max: 161

T:12 ( 2347) P:95 I:6500 C: 7692 Min: 12 Act: 19 Avg: 24 Max: 194

T:13 ( 2348) P:95 I:7000 C: 7142 Min: 12 Act: 64 Avg: 23 Max: 185

T:14 ( 2349) P:95 I:7500 C: 6666 Min: 12 Act: 20 Avg: 23 Max: 192

Now the thing is when getting the CPUs a little bit busy with something 
else(e.g. simultaneous running “while true; do echo loop; done” on in 
another shell) the same cyclictest results in:

root@host:[~]# cyclictest -c 0 -t15 -p 95 -i 500 -n -l 100000

policy: fifo: loadavg: 1.78 0.54 0.22 3/203 2431

T: 0 ( 2417) P:95 I:500 C: 100000 Min: 10 Act: 19 Avg: 16 Max: 44

T: 1 ( 2418) P:95 I:1000 C: 49998 Min: 10 Act: 16 Avg: 15 Max: 45

T: 2 ( 2419) P:95 I:1500 C: 33335 Min: 9 Act: 16 Avg: 15 Max: 33

T: 3 ( 2420) P:95 I:2000 C: 25000 Min: 10 Act: 14 Avg: 16 Max: 40

T: 4 ( 2421) P:95 I:2500 C: 20000 Min: 10 Act: 15 Avg: 16 Max: 33

T: 5 ( 2422) P:95 I:3000 C: 16667 Min: 10 Act: 15 Avg: 16 Max: 60

T: 6 ( 2423) P:95 I:3500 C: 14286 Min: 10 Act: 14 Avg: 16 Max: 35

T: 7 ( 2424) P:95 I:4000 C: 12500 Min: 12 Act: 24 Avg: 16 Max: 33

T: 8 ( 2425) P:95 I:4500 C: 11111 Min: 11 Act: 15 Avg: 16 Max: 34

T: 9 ( 2426) P:95 I:5000 C: 10000 Min: 10 Act: 15 Avg: 17 Max: 46

T:10 ( 2427) P:95 I:5500 C: 9091 Min: 10 Act: 15 Avg: 15 Max: 38

T:11 ( 2428) P:95 I:6000 C: 8333 Min: 10 Act: 13 Avg: 15 Max: 35

T:12 ( 2429) P:95 I:6500 C: 7692 Min: 11 Act: 13 Avg: 16 Max: 42

T:13 ( 2430) P:95 I:7000 C: 7143 Min: 11 Act: 12 Avg: 16 Max: 32

T:14 ( 2431) P:95 I:7500 C: 6666 Min: 11 Act: 15 Avg: 17 Max: 37

As can be seen the max-latency improved!

It’s like ordering in an empty restaurant. Takes hours to get the dinner 
if you are the only customer. As soon as some more people are around 
everything is getting faster...

So my guess is that the cpu goes into some sort of powersave mode from 
time to time. The wakeup from these powersave state results in 
latencies. Currently I am looking for a proper solution to this problem 
and what I have tried so far:

- BIOS upgrade
- Upgraded to linux-3.10.12-rt25
- Disabling all sort of powersave modes in the kernel config
- Activated cpufreq-performance (which seems to be ignored anyway)
- Searched the BIOS for any powersave modes to deactivate (none found)

I got rather clueless on what else to try! Does anyone has a helping 
hint on how to solve this one???

With many thanks in advance,

Stefan J.

--
To unsubscribe from this list: send the line "unsubscribe linux-rt-users" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: Difficulties with cyclictest on AMD CPU

[Clark Williams]

[-- Attachment #1: Type: text/plain, Size: 1028 bytes --]

On Sun, 27 Apr 2014 13:01:08 +0200
Stefan Jenisch <stefan.jenisch@gmail.com> wrote:
> 
> So my guess is that the cpu goes into some sort of powersave mode from 
> time to time. The wakeup from these powersave state results in 
> latencies. Currently I am looking for a proper solution to this problem 
> and what I have tried so far:
> 
> - BIOS upgrade
> - Upgraded to linux-3.10.12-rt25
> - Disabling all sort of powersave modes in the kernel config
> - Activated cpufreq-performance (which seems to be ignored anyway)
> - Searched the BIOS for any powersave modes to deactivate (none found)
> 
> I got rather clueless on what else to try! Does anyone has a helping 
> hint on how to solve this one???
> 
> With many thanks in advance,
> 
> Stefan J.
> 

Did you get this message when starting cyclictest?

# /dev/cpu_dma_latency set to 0us


That's a hack that is supposed to prevent the system from entering
deeper C-states than C-state 0. 

What version of rt-tests are you running?

Clark

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 198 bytes --]