[lkml]   [2020]   [Sep]   [28]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
Subject[PATCH v4 0/4] isolation: limit msix vectors to housekeeping CPUs
This is a follow-up posting for "[PATCH v3 0/4] isolation: limit msix vectors
to housekeeping CPUs".

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        

Proposed Fix
In this patch-set, the following changes are proposed:
- A generic API housekeeping_num_online_cpus() which is meant to return the
online housekeeping CPUs based on the hk_flag passed by the caller.
- 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 that returns the online housekeeping CPUs that are meant
to handle managed IRQ jobs.
- pci_alloc_irq_vector(): With the help of housekeeping_num_online_cpus() the
max_vecs passed in pci_alloc_irq_vector() is restricted only to the online
  housekeeping CPUs (designated for managed IRQ jobs) strictly 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.

- 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. Another way to verify is by checking
the number of IRQs that get created corresponding to a impacted device.
Similarly did a few more tests with different modes eg with only nohz_full,
isolcpus etc.

- 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
  + 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).

  # 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 v3[2]:
- Moved the logic to limit the max_vecs from pci_alloc_irq_vectors() to
pci_alloc_irq_vectors_affinity() as that's the exported interface and
drivers using this API also need to be fixed (suggestion from Bjorn Helgaas).

Changes from v2[3]:
- Renamed hk_num_online_cpus() with housekeeping_num_online_cpus() to keep
the naming convention consistent (based on a suggestion from Peter
Zijlstra and Frederic Weisbecker).
- Added an argument "enum hk_flags" to the housekeeping_num_online_cpus() API
to make it more usable in different use-cases (based on a suggestion from
Frederic Weisbecker).
- Replaced cpumask_weight(cpu_online_mask) with num_online_cpus() (suggestion
from Bjorn Helgaas).
- Modified patch commit messages and comment based on Bjorn Helgaas's

Changes from v1[4]:
- 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
- Since the hk_num_online_cpus() is self-explanatory, got rid of            
  the comment that was added previously.                                    
- 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.                                                                    


Nitesh Narayan Lal (4):
sched/isolation: API to get number of housekeeping CPUs
sched/isolation: Extend nohz_full to isolate managed IRQs
i40e: Limit msix vectors to housekeeping CPUs
PCI: Limit pci_alloc_irq_vectors() to housekeeping CPUs

drivers/net/ethernet/intel/i40e/i40e_main.c | 3 ++-
drivers/pci/msi.c | 18 ++++++++++++++++++
include/linux/sched/isolation.h | 9 +++++++++
kernel/sched/isolation.c | 2 +-
4 files changed, 30 insertions(+), 2 deletions(-)


 \ /
  Last update: 2020-09-28 20:36    [W:0.184 / U:9.480 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site