I've been working on this issue occasionally for a 2+ years now.
Unfortunately, it is not easy to solve and substantiated information on
it is also hard to come by.
My current, **empirically founded** solution to get interfering
interrupts down as much as possible is the following (last tested about
a year ago with kernel ~5.10):
1. Don't run anything on CPU 0. It's the boot CPU and may run some
Linux voodoo that cannot be transferred to other CPUs. Read this
somewhere, but haven't been able to substantiate that. Also, disable
SMT or keep the SMT-siblings isolated, too, to avoid contention on
cache and memory. May both be superstition, but I don't see any
potential down-side to it.
2. Use real-time priority (SCHED_FIFO w/ priority 99) for the DPDK
threads and
echo -1 > /proc/sys/kernel/sched_rt_runtime_us
to disable the runtime limit. With the runtime limit in place, the
SCHED_FIFO performance will be significantly worse than SCHED_OTHER.
3. Use the following kernel command line (similar to yours; isolates
all but 2 cores on a 12 core machine):
isolcpus=nohz,domain,2-11,14-23 nohz_full=2-11,14-23
rcu_nocbs=2-11,14-23 rcu_nocb_poll irqaffinity=0-1,12-13
Note that this only works properly with a NO_HZ kernel. The kernel
boot log should contain an error message if above parameters are
used on a non-NO_HZ kernel.
I compile my kernel with the following config enabled: NO_HZ_FULL,
RCU_NOCB_CPU, CONTEXT_TRACKING_FORCE
4. Force remaining interrupts off the isolated CPUs by first stopping
all of them and subsequently starting them again (some interrupts
may remain on the isolated CPUs regardless of irqaffinit=; mileage
may vary):
for CPU in $CPU_LIST ;
do echo 0 > /sys/devices/system/cpu/cpu$CPU/online
done
for CPU in $CPU_LIST ;
do echo 1 > /sys/devices/system/cpu/cpu$CPU/online
done
5. Check which interrupts still occur on the isolated CPUs while
running your DPDK progam via /proc/interrupts. I've had issues with
some hardware drivers' interrupts (e.g., RAID controllers) refusing
to be kicked off the isolated CPUs despite all of the above. Try to
move your sensitive threads to different CPUs.
6. Despair. Even if you succeed in getting all hardware interrupts
disabled, the kernel will still occasionally interrupt your program,
e.g., for some accounting business.
The last point is where intervention from the kernel side is required.
Work on that has been underway for several years [1,2,3], but nothing
has been mainlined yet.
Hope this helps. We've been able to get worst case packet forwarding
jitter down to less than 10 µs, with anything above 3 µs being very rare
(see attached histogram; your mileage may vary; measured by comparing
TSC values between DPDK rx and tx calls).
Best regards,
Carsten
[1] https://lwn.net/Articles/659490/
[2] https://lwn.net/Articles/816298/
[3]
https://lore.kernel.org/lkml/20220315153132.717153751@fedora.localdomain/
On 23.06.22 20:03, Antonio Di Bacco wrote:
I'm running a DPDK thread on an isolated core. I also set some flags
that could help keeping the core at rest on linux like: nosoftlockup
nohz_full rcu_nocbs irqaffinity.
Unfortunately the thread gets some interruptions that stop the thread
for about 20-30 micro seconds. This seems smal but my application
suffers a lot.
I also tried to use rte_thread_set_priority that indeed has a strong
effect but unfortunately creates problems to Linux (like network not
working).
Is there any other knob that could help running the DPDK thread with
minimum or no interruptions at all?
