Skip to contentSkip to conversation

Dynamic tick for 2.6.14 - what's the plan?

DISCUSSION Lee Revell 8 participants 18 messages Aug 29, 2005 – Sep 1, 2005

lore
Share & tools
lore

Lee Revell

author
#0 lore raw

The controversy over the introduction of CONFIG_HZ demonstrated the urgency of getting a dynamic tick solution merged before 2.6.14.

Anyone care to give a status report? Con, do you feel that the last version you posted is ready to go in?

Lee

#1 lore raw
Quoted 5 lines · Lee Revell

The controversy over the introduction of CONFIG_HZ demonstrated the urgency of getting a dynamic tick solution merged before 2.6.14.

Anyone care to give a status report? Con, do you feel that the last version you posted is ready to go in?

Last time I got interested in this, the management of the event queues was still a fairly major performance hit.

Has this overhead been brought down to reasonable levels?

Chris

#2 lore raw
Quoted 11 lines · Christopher Friesen
Quoted 5 lines · Lee Revell

The controversy over the introduction of CONFIG_HZ demonstrated the urgency of getting a dynamic tick solution merged before 2.6.14.

Anyone care to give a status report? Con, do you feel that the last version you posted is ready to go in?

Last time I got interested in this, the management of the event queues was still a fairly major performance hit.

Has this overhead been brought down to reasonable levels?

Srivatsa is still working on the smp-aware scalable version, so it's back in the development phase.

Cheers,
Con

#3 lore raw
Quoted 8 lines · Con Kolivas
Quoted 7 lines · Christopher Friesen
Quoted 5 lines · Lee Revell

The controversy over the introduction of CONFIG_HZ demonstrated the urgency of getting a dynamic tick solution merged before 2.6.14.

Anyone care to give a status report? Con, do you feel that the last version you posted is ready to go in?

Quoted 7 lines
Quoted 4 lines

Last time I got interested in this, the management of the event queues was still a fairly major performance hit.

Has this overhead been brought down to reasonable levels?

Srivatsa is still working on the smp-aware scalable version, so it's back in the development phase.

Has there been an updated version of Thomas's C-state bus-mastering measurement patch?

- Ted

#4 lore raw
Quoted 12 lines · Theodore Ts'o
Quoted 11 lines · Con Kolivas
Quoted 10 lines · Christopher Friesen
Quoted 5 lines · Lee Revell

The controversy over the introduction of CONFIG_HZ demonstrated the urgency of getting a dynamic tick solution merged before 2.6.14.

Anyone care to give a status report? Con, do you feel that the last version you posted is ready to go in?

Last time I got interested in this, the management of the event queues was still a fairly major performance hit.

Quoted 7 lines
Quoted 4 lines
Quoted 1 line

Has this overhead been brought down to reasonable levels?

Srivatsa is still working on the smp-aware scalable version, so it's back in the development phase.

Has there been an updated version of Thomas's C-state bus-mastering measurement patch?

Same issue, it's waiting on dynticks before being reworked.

Cheers,
Con

#5 lore raw
Quoted 15 lines · Con Kolivas
Quoted 14 lines · Theodore Ts'o
Quoted 13 lines · Con Kolivas
Quoted 11 lines · Christopher Friesen
Quoted 5 lines · Lee Revell

The controversy over the introduction of CONFIG_HZ demonstrated the urgency of getting a dynamic tick solution merged before 2.6.14.

Anyone care to give a status report? Con, do you feel that the last version you posted is ready to go in?

Last time I got interested in this, the management of the event queues was still a fairly major performance hit.

Has this overhead been brought down to reasonable levels?

Quoted 7 lines
Quoted 5 lines
Quoted 2 lines

Srivatsa is still working on the smp-aware scalable version, so it's back in the development phase.

Has there been an updated version of Thomas's C-state bus-mastering measurement patch?

Same issue, it's waiting on dynticks before being reworked.

Also one more minor issue; Dyntick can cause slow boots with dyntick enabled from boot because the there's not much in the timer queue until init.

This probably does not show up much on x86 though because of the short hardware timers.

Tony

#6 lore raw

On Tuesday 30 August 2005 13:31, Tony Lindgren wrote:
[snip]

Quoted 9 lines
Quoted 2 lines

Same issue, it's waiting on dynticks before being reworked.

Also one more minor issue; Dyntick can cause slow boots with dyntick enabled from boot because the there's not much in the timer queue until init.

This probably does not show up much on x86 though because of the short hardware timers.

You could disable it until jiffies >= 0; this covers the boot criteria and still allows for moderate savings post boot (though maybe on embedded systems the delay is too long?).

Signature · 6 lines
Cheers,
Alistair.

'No sense being pessimistic, it probably wouldn't work anyway.'
Third year Computer Science undergraduate.
1F2 55 South Clerk Street, Edinburgh, UK.
#7 lore raw
Quoted 9 lines · Alistair John Strachan

On Tuesday 30 August 2005 13:31, Tony Lindgren wrote:
[snip]

Quoted 7 lines
Quoted 2 lines

Same issue, it's waiting on dynticks before being reworked.

Also one more minor issue; Dyntick can cause slow boots with dyntick enabled from boot because the there's not much in the timer queue until init.

Quoted 6 lines
Quoted 2 lines

This probably does not show up much on x86 though because of the short hardware timers.

You could disable it until jiffies >= 0; this covers the boot criteria and still allows for moderate savings post boot (though maybe on embedded systems the delay is too long?).

Yeah, that's true. Or just enable it from an init script via sysfs.

Tony

#8 lore raw
Quoted 10 lines · Tony Lindgren
Quoted 9 lines · Alistair John Strachan

On Tuesday 30 August 2005 13:31, Tony Lindgren wrote:
[snip]

Quoted 7 lines
Quoted 2 lines

Same issue, it's waiting on dynticks before being reworked.

Also one more minor issue; Dyntick can cause slow boots with dyntick enabled from boot because the there's not much in the timer queue until init.

Quoted 8 lines
Quoted 6 lines
Quoted 2 lines

This probably does not show up much on x86 though because of the short hardware timers.

You could disable it until jiffies >= 0; this covers the boot criteria and still allows for moderate savings post boot (though maybe on embedded systems the delay is too long?).

Yeah, that's true. Or just enable it from an init script via sysfs.

ehh
why does it cause slow boots?
if that kind of behavior changes... isn't that a sign there is a fundamental bug still ?

#9 lore raw
Quoted 18 lines · Arjan van de Ven
Quoted 17 lines · Tony Lindgren
Quoted 15 lines · Alistair John Strachan

On Tuesday 30 August 2005 13:31, Tony Lindgren wrote:
[snip]

Quoted 9 lines
Quoted 2 lines

Same issue, it's waiting on dynticks before being reworked.

Also one more minor issue; Dyntick can cause slow boots with dyntick enabled from boot because the there's not much in the timer queue until init.

This probably does not show up much on x86 though because of the short hardware timers.

You could disable it until jiffies >= 0; this covers the boot criteria and still allows for moderate savings post boot (though maybe on embedded systems the delay is too long?).

Quoted 6 lines
Quoted 1 line

Yeah, that's true. Or just enable it from an init script via sysfs.

ehh
why does it cause slow boots?
if that kind of behavior changes... isn't that a sign there is a fundamental bug still ?

Well it seems like the next_timer_interrupt is something like 400 jiffies away and RCU code waits for completion for example in the network code.

Tony

#10 lore raw
Quoted 3 lines · Tony Lindgren

Well it seems like the next_timer_interrupt is something like 400 jiffies away and RCU code waits for completion for example in the network code.

I had a patch to fix the problem of "RCU grace period extended because of sleeping idle CPUs". I had posted the patch here:

http://marc.theaimsgroup.com/?l=linux-kernel&m=111556608901657&w=2

Will send out this patch against latest tree for Andrew to pick it.

Signature · 5 lines
Thanks and Regards,
Srivatsa Vaddagiri,
Linux Technology Center,
IBM Software Labs,
Bangalore, INDIA - 560017
#11 lore raw
In reply to Tony Lindgren
Quoted 9 lines
Quoted 5 lines

ehh
why does it cause slow boots?
if that kind of behavior changes... isn't that a sign there is a fundamental bug still ?

Well it seems like the next_timer_interrupt is something like 400 jiffies away and RCU code waits for completion for example in the network code.

that sounds like a fundamental issue that really needs to be fixed first!

#12 lore raw
Quoted 2 lines · Arjan van de Ven

that sounds like a fundamental issue that really needs to be fixed first!

It should be fixed by the patch here:
http://marc.theaimsgroup.com/?l=linux-kernel&m=111556608901657&w=2

Tony,
I don't see any slow bootups on x86 because of dyn-tick.
If you still see them in your env, can you test with the patch above?

Recovering time after sleep is the single biggest problem that I seem to have, even while using ACPI PM timer (forget TSC). Time can drift by couple of seconds after few hours. I have made some changes to the lost tick calculation in timer_pm.c after which it seems to be stable on some machines, but I cant repeat that on other (maybe newer) machines. Will post out all the changes I have pretty soon.

Signature · 5 lines
Thanks and Regards,
Srivatsa Vaddagiri,
Linux Technology Center,
IBM Software Labs,
Bangalore, INDIA - 560017
#13 lore raw
Quoted 6 lines · Srivatsa Vaddagiri
Quoted 2 lines · Arjan van de Ven

that sounds like a fundamental issue that really needs to be fixed first!

It should be fixed by the patch here:
http://marc.theaimsgroup.com/?l=linux-kernel&m=111556608901657&w=2

Actually, a solution to take care of sleeping CPUs was there quite some time back. The above patch only fixes a race in that solution.

Tony,
Which kernel version did you see the slow bootup?

Signature · 5 lines
Thanks and Regards,
Srivatsa Vaddagiri,
Linux Technology Center,
IBM Software Labs,
Bangalore, INDIA - 560017
#14 lore raw
Quoted 5 lines · Srivatsa Vaddagiri
Quoted 4 lines · Srivatsa Vaddagiri
Quoted 2 lines · Arjan van de Ven

that sounds like a fundamental issue that really needs to be fixed first!

Quoted 8 lines
Quoted 2 lines

It should be fixed by the patch here:
http://marc.theaimsgroup.com/?l=linux-kernel&m=111556608901657&w=2

Actually, a solution to take care of sleeping CPUs was there quite some time back. The above patch only fixes a race in that solution.

Tony,
Which kernel version did you see the slow bootup?

I'll try it out, but sounds like it won't help then.

This is on ARM OMAP and 2.6.12. I haven't been able to play with the x86 stuff lately, hopefully will have a chance soonish...

Tony

#15 lore raw
Quoted 12 lines · Tony Lindgren
Quoted 11 lines · Srivatsa Vaddagiri
Quoted 6 lines · Srivatsa Vaddagiri
Quoted 2 lines · Arjan van de Ven

that sounds like a fundamental issue that really needs to be fixed first!

It should be fixed by the patch here:
http://marc.theaimsgroup.com/?l=linux-kernel&m=111556608901657&w=2

Actually, a solution to take care of sleeping CPUs was there quite some time back. The above patch only fixes a race in that solution.

Quoted 7 lines
Quoted 2 lines

Tony,
Which kernel version did you see the slow bootup?

I'll try it out, but sounds like it won't help then.

This is on ARM OMAP and 2.6.12. I haven't been able to play with the x86 stuff lately, hopefully will have a chance soonish...

It could be that the slow bootup is currently just OMAP specific issue, I'll post more info when I find out.

Tony

#16 lore raw
In reply to Christopher Friesen
Quoted 4 lines · Christopher Friesen

Last time I got interested in this, the management of the event queues was still a fairly major performance hit.

Has this overhead been brought down to reasonable levels?

Hmm ..I dont see any event queues being managed by dyn-tick patch. Are you referring to some old version which I havent seen perhaps? If so, what were those event queues used for?

Signature · 5 lines
Thanks and Regards,
Srivatsa Vaddagiri,
Linux Technology Center,
IBM Software Labs,
Bangalore, INDIA - 560017
#17 lore raw
Quoted 5 lines · Srivatsa Vaddagiri
Quoted 1 line · Christopher Friesen

Last time I got interested in this, the management of the event queues was still a fairly major performance hit.

Hmm ..I dont see any event queues being managed by dyn-tick patch. Are you referring to some old version which I havent seen perhaps? If so, what were those event queues used for?

Oh, sorry. I think I got this confused with the tickless patch.

Chris