Skip to contentSkip to conversation

cpufreq affects traffic control rates

DISCUSSION DervishD 3 participants 5 messages Aug 28, 2007 – Aug 29, 2007

lore
Share & tools
lore

DervishD

author
#0 lore raw

Hi all :)

I noticed lately that my traffic control rates were being very slow,
about 40% less than expected, and finally spotted the problem: cpufreq.

Looks like HTB puts buckets according to the requested rate but
assuming that the CPU is running at its default clock or something like that.

Any way of fixing this without disabling cpufreq?

I'm using kernel 2.6.20.14, Athlon64 1000/1800MHz, HZ=1000 and a
combination of HTB/SFQ in my traffic control.

Thanks a lot in advance :)

Raúl Núñez de Arenas Coronado

Signature · 2 lines
Linux Registered User 88736 | http://www.dervishd.net
It's my PC and I'll cry if I want to... RAmen!
#1 lore raw
Quoted 9 lines · On Tue, 28 Aug 2007 09:51:55 +0200 DervishD <[email protected]> wrote:

Hi all :)

I noticed lately that my traffic control rates were being very slow,
about 40% less than expected, and finally spotted the problem: cpufreq.

Looks like HTB puts buckets according to the requested rate but
assuming that the CPU is running at its default clock or something like that.

Quoted 9 lines

Any way of fixing this without disabling cpufreq?

I'm using kernel 2.6.20.14, Athlon64 1000/1800MHz, HZ=1000 and a
combination of HTB/SFQ in my traffic control.

Thanks a lot in advance :)

Raúl Núñez de Arenas Coronado

Is the problem configuration of network scheduler clock? In 2.6.20 and earlier, you could use CPU cycle counter (later kernels only use time of day). So try switching to jiffies or gettimeofday.

Signature · 1 line
Stephen Hemminger <[email protected]>
author
#2 lore raw

Hi Stephen :)

* Stephen Hemminger <[email protected]> dixit:

Quoted 9 lines
Quoted 3 lines · On Tue, 28 Aug 2007 09:51:55 +0200 DervishD <[email protected]> wrote:

I noticed lately that my traffic control rates were being very slow, about 40% less than expected, and finally spotted the
problem: cpufreq.

Is the problem configuration of network scheduler clock? In 2.6.20 and earlier, you could use CPU cycle counter (later kernels only use time of day). So try switching to jiffies or gettimeofday.

Yes, the problem is that I'm stupid and I thought I was using
jiffies, but I have CONFIG_NET_SCH_CLK_CPU. I'll rebuild my kernel ASAP and set the proper value...

Thanks for your help. It's embarrassing I didn't check the clock
source for the packet scheduler...

Sorry for the noise.

Raúl Núñez de Arenas Coronado

Signature · 2 lines
Linux Registered User 88736 | http://www.dervishd.net
It's my PC and I'll cry if I want to... RAmen!
#3 lore raw
In reply to Stephen Hemminger
Quoted 10 lines · Stephen Hemminger
Quoted 7 lines

I noticed lately that my traffic control rates were being very slow,
about 40% less than expected, and finally spotted the problem: cpufreq.

Looks like HTB puts buckets according to the requested rate but
assuming that the CPU is running at its default clock or something like that.

Is the problem configuration of network scheduler clock? In 2.6.20 and earlier, you could use CPU cycle counter (later kernels only use time of day). So try switching to jiffies or gettimeofday.

This should not have been removed. CPUs with constant_tsc could still be used with "CPU cycle counter" method.

Jan

#4 lore raw
Quoted 13 lines · On Wed, 29 Aug 2007 17:26:24 +0200 (CEST) Jan Engelhardt…
Quoted 10 lines · Stephen Hemminger
Quoted 7 lines

I noticed lately that my traffic control rates were being very slow,
about 40% less than expected, and finally spotted the problem: cpufreq.

Looks like HTB puts buckets according to the requested rate but
assuming that the CPU is running at its default clock or something like that.

Is the problem configuration of network scheduler clock? In 2.6.20 and earlier, you could use CPU cycle counter (later kernels only use time of day). So try switching to jiffies or gettimeofday.

Quoted 5 lines

This should not have been removed. CPUs with constant_tsc could still be used with "CPU cycle counter" method.

Jan

It was removed because the core clock subsystem now manages the choice of CPU clock correctly and should provide the highest resolution possible on the hardware.

Plus a whole bunch of grotty macro code gets removed. And the math handling was more standard/simpler.

Signature · 1 line
Stephen Hemminger <[email protected]>