Re: [RFC v2] timers: Don't wake ktimersoftd on every tick

From: Haris Okanovic
Date: Wed Dec 28 2016 - 15:40:02 EST


On 12/23/2016 11:28 AM, Sebastian Andrzej Siewior wrote:
On 2016-12-13 15:44:05 [-0600], Haris Okanovic wrote:
Changed the way timers are collected per Julia and Thomas'

I can only see Julia's response to the initial thread.


I should have been more clear. Thomas commented on irc and recommended Julia's approach.

recommendation: Expired timers are now collected in interrupt context
and fired in ktimersoftd to avoid double-walk of `pending_map`.

This is implemented by storing lists of expired timers in timer_base,
which carries a memory overhead 9*sizeof(pointer) per CPU. The timer
system uses hlist's which don't have end-node references, making it
impossible to merge 2 hlist's in constant time. I.e. Merging requires
walking one list. I also considered switching `vectors` to regular
list's which don't have this limitations, but that approach has the same
memory overhead. list_head is bigger than hlist_head by sizeof(pointer)
and is instantiated 9+ times per CPU as `vectors`. I believe the only
way to trim overhead is to spend more CPU cycles in interrupt context
either in list merging (unbounded operation) or the original double-walk
implementation. Any suggestions/preferences?

As before, a 6h run of cyclictest without CPU affinity shows decrease in
22-70us latency range.
what does this mean? Your cyclictest runs on a random CPU with one thread
only?


Yes. My point is that cyclictest only shows a significant difference (before and after this change) when `-S` is not used.

No change in max jitter.
Does this mean your average latency went down 20-70us and your max is
the same?


Yes. Average latency (20-70us range) goes down in a single-threaded run of cyclictest. Max jitter stays the same in both single and multi-thread runs.

No change when `-S` is
used.

-S gives you one thread per core, makes sure it stays on that core and
uses clock_nanosleep().

clock_nanosleep() should be used no matter what.


[Before/after traces]

ftp://ftp.ni.com/outgoing/tp02-timer-peek-traces.tgz
(Email me if link dies. Server periodically purges old files.)

[Hardware/software/config]

NI cRIO-9033
2 core Intel Atom CPU

Kernel 4.8.6-rt5
CONFIG_HZ_PERIODIC=y

[Outstanding concerns/issues/questions]

I'm relatively new to the timer subsystem, so please feel free to poke
as many holes as possible in this change. A few things that concern me
at the moment are:

Can jiffies change while one or more cpus is inside tick_sched_timer(),
in interrupt context? I'm copying jiffies to a local variable in
find_expired_timers() to ensure it doesn't run unbounded, but I'm not
sure if that's necessary.

It could change. Only the house keeping does update jiffies in
tick_sched_do_timer().

Any special considerations for testing NO_HZ builds? (Other than letting
it run idle for a while)

timers_dead_cpu() presently asserts no timer callback is actively
running, which suggests that timers must be canceled prior to disabling
CPUs; otherwise, there's a race between active timers and hotplug
which can crash the whole kernel. Is this a safe assumption to make and
are there any special considerations for CPU hotplug testing?

timers_dead_cpu() and hrtimers_dead_cpu() migrate timer away. At that
point the CPU should be down already so a timer can't run on that CPU.

Other tests/performance benchmark I should run?

Source: https://github.com/harisokanovic/linux/tree/dev/hokanovi/timer-peek-v2

Thanks,
Haris

Sebastian


-- Haris