After enabling CONFIG_IRQ_TIME_ACCOUNTING to track IRQ pressure in our
container environment, we encountered several user-visible behavioral
changes:
- Interrupted IRQ/softirq time is excluded in the cpuacct cgroup
This breaks userspace applications that rely on CPU usage data from
cgroups to monitor CPU pressure. This patchset resolves the issue by
ensuring that IRQ/softirq time is included in the cgroup of the
interrupted tasks.
- getrusage(2) does not include time interrupted by IRQ/softirq
Some services use getrusage(2) to check if workloads are experiencing CPU
pressure. Since IRQ/softirq time is no longer included in task runtime,
getrusage(2) can no longer reflect the CPU pressure caused by heavy
interrupts.
This patchset addresses the first issue, which is relatively
straightforward. Once this solution is accepted, I will address the second
issue in a follow-up patchset.
Enabling CONFIG_IRQ_TIME_ACCOUNTING results in the CPU
utilization metric excluding the time spent in IRQs. This means we
lose visibility into how long the CPU was actually interrupted in
comparison to its total utilization. The user will misleadlingly
consider the *interrupted irq time* as *sleep time* as follows,
|<----Runtime---->|<----Sleep---->|<----Runtime---->|<---Sleep-->|
While actually it is:
|<----Runtime---->|<--Interrupted time-->|<----Runtime---->|<---Sleep-->|
Currently, the only ways to monitor interrupt time are through IRQ PSI or
the IRQ time recorded in delay accounting. However, these metrics are
independent of CPU utilization, which makes it difficult to combine them
into a single, unified measure
CPU utilization is a critical metric for almost all workloads, and
it's problematic if it fails to reflect the full extent of system
pressure. This situation is similar to iowait: when a task is in
iowait, it could be due to other tasks performing I/O. It doesn’t
matter if the I/O is being done by one of your tasks or by someone
else's; what matters is that your task is stalled and waiting on I/O.
Similarly, a comprehensive CPU utilization metric should reflect all
sources of pressure, including IRQ time, to provide a more accurate
representation of workload behavior.
One of the applications impacted by this issue is our Redis load-balancing
service. The setup operates as follows:
----------------
| Load Balancer|
----------------
/ | | \
/ | | \
Server1 Server2 Server3 ... ServerN
Although the load balancer's algorithm is complex, it follows some core
principles:
- When server CPU utilization increases, it adds more servers and deploys
additional instances to meet SLA requirements.
- When server CPU utilization decreases, it scales down by decommissioning
servers and reducing the number of instances to save on costs.
On our servers, the majority of IRQ/softIRQ activity originates from
network traffic, and we consistently enable Receive Flow Steering
(RFS) [0]. This configuration ensures that softIRQs are more likely to
interrupt the tasks responsible for processing the corresponding
packets. As a result, the distribution of softIRQs is not random but
instead closely aligned with the packet-handling tasks.
The load balancer is malfunctioning due to the exclusion of IRQ time from
CPU utilization calculations. What's worse, there is no effective way to
add the irq time back into the CPU utilization based on current
available metrics. Therefore, we have to change the kernel code.
Link: https://lwn.net/Articles/381955/ [0]
Changes:
v7->v8:
- Fix a build failure report by kernel test robot
v6->v7: https://lore.kernel.org/all/20241215032315.43698-1-laoar.shao@xxxxxxxxx/
- Fix psi_show() (Michal)
v5->v6: https://lore.kernel.org/all/20241211131729.43996-1-laoar.shao@xxxxxxxxx/
- Return EOPNOTSUPP in psi_show() if irqtime is disabled (Michal)
- Collect Reviewed-by from Michal
v4->v5: https://lore.kernel.org/all/20241108132904.6932-1-laoar.shao@xxxxxxxxx/
- Don't use static key in the IRQ_TIME_ACCOUNTING=n case (Peter)
- Rename psi_irq_time to irq_time (Peter)
- Use CPUTIME_IRQ instead of CPUTIME_SOFTIRQ (Peter)
v3->v4: https://lore.kernel.org/all/20241101031750.1471-1-laoar.shao@xxxxxxxxx/
- Rebase
v2->v3:
- Add a helper account_irqtime() to avoid redundant code (Johannes)
v1->v2: https://lore.kernel.org/cgroups/20241008061951.3980-1-laoar.shao@xxxxxxxxx/
- Fix lockdep issues reported by kernel test robot <oliver.sang@xxxxxxxxx>
v1: https://lore.kernel.org/all/20240923090028.16368-1-laoar.shao@xxxxxxxxx/
Yafang Shao (4):
sched: Define sched_clock_irqtime as static key
sched: Don't account irq time if sched_clock_irqtime is disabled
sched, psi: Don't account irq time if sched_clock_irqtime is disabled
sched: Fix cgroup irq time for CONFIG_IRQ_TIME_ACCOUNTING
kernel/sched/core.c | 77 +++++++++++++++++++++++++++++-------------
kernel/sched/cputime.c | 16 ++++-----
kernel/sched/psi.c | 16 ++++-----
kernel/sched/sched.h | 15 +++++++-
kernel/sched/stats.h | 7 ++--
5 files changed, 86 insertions(+), 45 deletions(-)
--
2.43.5
