This is a follow-up posting for "[RFC v1 0/3] isolation: limit msix vectors
based on housekeeping CPUs".
Issue
=====
With the current implementation device drivers while creating their MSIX
vectors only take num_online_cpus() into consideration which works quite well
for a non-RT environment, but in an RT environment that has a large number of
isolated CPUs and very few housekeeping CPUs this could lead to a problem.
The problem will be triggered when something like tuned will try to move all
the IRQs from isolated CPUs to the limited number of housekeeping CPUs to
prevent interruptions for a latency-sensitive workload that will be running on
the isolated CPUs. This failure is caused because of the per CPU vector
limitation.
Proposed Fix
============
In this patch-set, the following changes are proposed:
- A generic API hk_num_online_cpus() which is meant to return the online
housekeeping CPUs that are meant to handle managed IRQ jobs.
- i40e: Specifically for the i40e driver the num_online_cpus() used in
i40e_init_msix() to calculate numbers msix vectors is replaced with the above
defined API. This is done to restrict the number of msix vectors for i40e in
RT environments.
- pci_alloc_irq_vector(): With the help of hk_num_online_cpus() the max_vecs
passed in pci_alloc_irq_vector() is restricted only to the online
housekeeping CPUs only in an RT environment. However, if the min_vecs exceeds
the online housekeeping CPUs, max_vecs is limited based on the min_vecs
instead.
Future Work
===========
- In the previous upstream discussion [1], it was decided that it would be
better if we can have a generic framework that can be consumed by all the
drivers to fix this kind of issue. However, it will be a long term work,
and since there are RT workloads that are getting impacted by the reported
issue. We agreed upon the proposed per-device approach for now.
Testing
=======
Functionality:
- To test that the issue is resolved with i40e change I added a tracepoint
in i40e_init_msix() to find the number of CPUs derived for vector creation
with and without tuned's realtime-virtual-host profile. As per expectation
with the profile applied I was only getting the number of housekeeping CPUs
and all available CPUs without it.
Similarly did a few more tests with different modes eg with only
nohz_full, isolcpus etc.
Performance:
- To analyze the performance impact I have targetted the change introduced in
pci_alloc_irq_vectors() and compared the results against a vanilla kernel
(5.9.0-rc3) results.
Setup Information:
+ I had a couple of 24-core machines connected back to back via a couple of
mlx5 NICs and I analyzed the average bitrate for server-client TCP and UDP
transmission via iperf.
+ To minimize the Bitrate variation of iperf TCP and UDP stream test I have
applied the tuned's network-throughput profile and disabled HT.
Test Information:
+ For the environment that had no isolated CPUs:
I have tested with single stream and 24 streams (same as that of online
CPUs).
+ For the environment that had 20 isolated CPUs:
I have tested with single stream, 4 streams (same as that the number of
housekeeping) and 24 streams (same as that of online CPUs).
Results:
# UDP Stream Test:
+ There was no degradation observed in UDP stream tests in both
environments. (With isolated CPUs and without isolated CPUs after the
introduction of the patches).
# TCP Stream Test - No isolated CPUs:
+ No noticeable degradation was observed.
# TCP Stream Test - With isolated CPUs:
+ Multiple Stream (4) - Average degradation of around 5-6%
+ Multiple Stream (24) - Average degradation of around 2-3%
+ Single Stream - Even on a vanilla kernel the Bitrate observed for
a TCP single stream test seem to vary
significantly across different runs (eg. the %
variation between the best and the worst case on
a vanilla kernel was around 8-10%). A similar
variation was observed with the kernel that
included my patches. No additional degradation
was observed.
If there are any suggestions for more performance evaluation, I would
be happy to discuss/perform them.
Changes from v1[2]:
==================
Patch1:
- Replaced num_houskeeeping_cpus() with hk_num_online_cpus() and started using
the cpumask corresponding to HK_FLAG_MANAGED_IRQ to derive the number of
online housekeeping CPUs. This is based on Frederic Weisbecker's suggestion.
- Since the hk_num_online_cpus() is self-explanatory, got rid of
the comment that was added previously.
Patch2:
- Added a new patch that is meant to enable managed IRQ isolation for nohz_full
CPUs. This is based on Frederic Weisbecker's suggestion.
Patch4 (PCI):
- For cases where the min_vecs exceeds the online housekeeping CPUs, instead
of skipping modification to max_vecs, started restricting it based on the
min_vecs. This is based on a suggestion from Marcelo Tosatti.
[1] https://lore.kernel.org/lkml/20200922095440.GA5217@lenoir/
[2] https://lore.kernel.org/lkml/[email protected]/
Nitesh Narayan Lal (4):
sched/isolation: API to get housekeeping online CPUs
sched/isolation: Extend nohz_full to isolate managed IRQs
i40e: limit msix vectors based on housekeeping CPUs
PCI: Limit pci_alloc_irq_vectors as per housekeeping CPUs
drivers/net/ethernet/intel/i40e/i40e_main.c | 3 ++-
include/linux/pci.h | 15 +++++++++++++++
include/linux/sched/isolation.h | 13 +++++++++++++
kernel/sched/isolation.c | 2 +-
4 files changed, 31 insertions(+), 2 deletions(-)
--