On 2024/8/15 04:58, Mateusz Guzik wrote:
On Wed, Aug 14, 2024 at 6:15 PM Wen Yang <wen.yang@xxxxxxxxx> wrote:
On 2024/8/11 18:26, Mateusz Guzik wrote:
On Sun, Aug 11, 2024 at 04:59:54PM +0800, Wen Yang wrote:
For the NON-SEMAPHORE eventfd, a write (2) call adds the 8-byte integer
value provided in its buffer to the counter, while a read (2) returns the
8-byte value containing the value and resetting the counter value to 0.
Therefore, the accumulated value of multiple writes can be retrieved by a
single read.
However, the current situation is to immediately wake up the read thread
after writing the NON-SEMAPHORE eventfd, which increases unnecessary CPU
overhead. By introducing a configurable rate limiting mechanism in
eventfd_write, these unnecessary wake-up operations are reduced.
[snip]
# ./a.out -p 2 -s 3
The original cpu usage is as follows:
09:53:38 PM CPU %usr %nice %sys %iowait %irq %soft %steal %guest %gnice %idle
09:53:40 PM 2 47.26 0.00 52.74 0.00 0.00 0.00 0.00 0.00 0.00 0.00
09:53:40 PM 3 44.72 0.00 55.28 0.00 0.00 0.00 0.00 0.00 0.00 0.00
09:53:40 PM CPU %usr %nice %sys %iowait %irq %soft %steal %guest %gnice %idle
09:53:42 PM 2 45.73 0.00 54.27 0.00 0.00 0.00 0.00 0.00 0.00 0.00
09:53:42 PM 3 46.00 0.00 54.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
09:53:42 PM CPU %usr %nice %sys %iowait %irq %soft %steal %guest %gnice %idle
09:53:44 PM 2 48.00 0.00 52.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
09:53:44 PM 3 45.50 0.00 54.50 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Then enable the ratelimited wakeup, eg:
# ./a.out -p 2 -s 3 -r1000 -c2
Observing a decrease of over 20% in CPU utilization (CPU # 3, 54% ->30%), as shown below:
10:02:32 PM CPU %usr %nice %sys %iowait %irq %soft %steal %guest %gnice %idle
10:02:34 PM 2 53.00 0.00 47.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
10:02:34 PM 3 30.81 0.00 30.81 0.00 0.00 0.00 0.00 0.00 0.00 38.38
10:02:34 PM CPU %usr %nice %sys %iowait %irq %soft %steal %guest %gnice %idle
10:02:36 PM 2 48.50 0.00 51.50 0.00 0.00 0.00 0.00 0.00 0.00 0.00
10:02:36 PM 3 30.20 0.00 30.69 0.00 0.00 0.00 0.00 0.00 0.00 39.11
10:02:36 PM CPU %usr %nice %sys %iowait %irq %soft %steal %guest %gnice %idle
10:02:38 PM 2 45.00 0.00 55.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
10:02:38 PM 3 27.08 0.00 30.21 0.00 0.00 0.00 0.00 0.00 0.00 42.71
Where are these stats from? Is this from your actual program you coded
the feature for?
The program you inlined here does next to nothing in userspace and
unsurprisingly the entire thing is dominated by kernel time, regardless
of what event rate can be achieved.
For example I got: /a.out -p 2 -s 3 5.34s user 60.85s system 99% cpu 66.19s (1:06.19) total
Even so, looking at perf top shows me that a significant chunk is
contention stemming from calls to poll -- perhaps the overhead will
sufficiently go down if you epoll instead?
We have two threads here, one publishing and one subscribing, running on
CPUs 2 and 3 respectively. If we further refine and collect performance
data on CPU 2, we will find that a large amount of CPU is consumed on
the spin lock of the wake-up logic of event write, for example:
# perf top -C 2 -e cycles:k
65.80% [kernel] [k] do_syscall_64
14.71% [kernel] [k] _raw_spin_unlock_irq
7.54% [kernel] [k] __fget_light
4.52% [kernel] [k] ksys_write
1.94% [kernel] [k] vfs_write
1.43% [kernel] [k] _copy_from_user
0.87% [kernel] [k] common_file_perm
0.61% [kernel] [k] aa_file_perm
0.46% [kernel] [k] eventfd_write
One of its call stacks:
|--6.39%--vfs_write
| --5.46%--eventfd_write
| --4.73%--_raw_spin_unlock_irq
> I think the idea is pretty dodgey. If the consumer program can tolerate
some delay in event processing, this probably can be massaged entirely in
userspace.
If your real program has the wake up rate so high that it constitutes a
tangible problem I wonder if eventfd is even the right primitive to use
-- perhaps something built around shared memory and futexes would do the
trick significantly better?
Thank you for your feedback.
This demo comes from the real world: the test vehicle has sensors with
multiple cycles (such as 1ms, 5ms, 10ms, etc.), and due to the large
number of sensors, data is reported at all times. The publisher reported
data through libzmq and went to the write logic of eventfd, frequently
waking up the receiver. We collected flame graph and observed that a
significant amount of CPU was consumed in this path: eventfd_write ->
_raw_spin_unlock_irq.
We did modify a lot of code in user mode on the test vehicle to avoid
this issue, such as not using wake-up, not using eventfd, the publisher
writing shared memory directly, the receiver periodically extracting the
content of shared memory, and so on.
Well I don't have the full picture and whatnot, but given the
additional info you posted here I even more strongly suspect eventfd
is a bad fit. AFAICS this boils down to batching a number of updates
and collecting them at some interval.
With the assumption that updates to the eventfd counter are guaranteed
to not overflow within the wakeup delay and that there is constant
traffic, I'm suspect you would get the expected speed up by using
timerfd to wake the consumer up periodically. Then you would only
issue an eventfd read when the timerfd tells you time is up. You would
(e)poll only on that as well, never on the eventfd.
Even so, as is I think this wants a page shared between producer(s)
and the consumer updating everything with atomics and the consumer
collecting it periodically (atomic add on one side, atomic swap with 0
on the consumer, I don't know the c11 intrinsics). It would be
drastically cheaper all around.
Thank you for your suggestion.
By using these methods above instead of eventfd, CPU consumption can
indeed be reduced.
But this requires modifying some user mode programs. Some of the
programs on the test vehicle are our own and can be modified; But there
is still a portion from various suppliers, and some even only deliver
binary, which is difficult to change.
And the kernel is open source, if it can be optimized, all user mode
programs can benefit from it.
You also mentioned that "AFAICS this boils down to batching a number of
updates and collecting them at some interval."
Yes, it's also similar to 'TCP's silly windw syndrome':
eventfd_write has two stages: adding one to the counter and waking up
the reader process. The former has low overhead, while the latter has
very high overhead. And the current situation is:
Every time the counter is incremented by 1, the read side process needs
to be awakened. When such operations are frequently performed, a lot of
time is wasted on awakening.
This patch is also inspired by algorithms such as Nagle and Cork. It
attempts to delay wake-up, accumulate a larger counter value, and then
wake up the reader process to consume the accumulated counter value at once.
Eventfd has already provided the NON-SEMAPHORE attribute, but it has not
been used yet. We look forward to your collaboration in using it
together to solve such problems.
--
Best wishes,
Wen
