kernel/linux
0
Fork 0
mirror of https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/ synced 2026-10-14 05:55:17 +00:00
Code Issues Releases Packages Activity

mlx5-updates-2021-06-14

Browse source 
1) Trivial Lag refactroing in preparation for upcomming Single FDB lag feature
  - First 3 patches
 
 2) Scalable IRQ distriburion for Sub-functions
 
 A subfunction (SF) is a lightweight function that has a parent PCI
 function (PF) on which it is deployed.
 
 Currently, mlx5 subfunction is sharing the IRQs (MSI-X) with their
 parent PCI function.
 
 Before this series the PF allocates enough IRQs to cover
 all the cores in a system, Newly created SFs will re-use all the IRQs
 that the PF has allocated for itself.
 Hence, the more SFs are created, there are more EQs per IRQs. Therefore,
 whenever we handle an interrupt, we need to pull all SFs EQs and PF EQs
 instead of PF EQs without SFs on the system. This leads to a hard impact
 on the performance of SFs and PF.
 
 For example, on machine with:
 Intel(R) Xeon(R) CPU E5-2697 v3 @ 2.60GHz with 56 cores.
 PCI Express 3 with BW of 126 Gb/s.
 ConnectX-5 Ex; EDR IB (100Gb/s) and 100GbE; dual-port QSFP28; PCIe4.0 x16.
 
 test case: iperf TX BW single CPU, affinity of app and IRQ are the same.
 PF only: no SFs on the system, 56 IRQs.
 SF (before), 250 SFs Sharing the same 56 IRQs .
 SF (now),    250 SFs + 255 avaiable IRQs for the NIC. (please see IRQ spread scheme below).
 
 	    application SF-IRQ  channel   BW(Gb/sec)         interrupts/sec
             iperf TX            affinity
 PF only     cpu={0}     cpu={0} cpu={0}   79                 8200
 SF (before) cpu={0}     cpu={0} cpu={0}   51.3 (-35%)        9500
 SF (now)    cpu={0}     cpu={0} cpu={0}   78 (-2%)           8200
 
 command:
 $ taskset -c 0 iperf -c 11.1.1.1 -P 3 -i 6 -t 30 | grep SUM
 
 The different between the SF examples is that before this series we
 allocate num_cpus (56) IRQs, and all of them were shared among the PF
 and the SFs. And after this series, we allocate 255 IRQs, and we spread
 the SFs among the above IRQs. This have significantly decreased the load
 on each IRQ and the number of EQs per IRQ is down by 95% (251->11).
 
 In this patchset the solution proposed is to have a dedicated IRQ pool
 for SFs to use. the pool will allocate a large number of IRQs
 for SFs to grab from in order to minimize irq sharing between the
 different SFs.
 IRQs will not be requested from the OS until they are 1st requested by
 an SF consumer, and will be eventually released when the last SF consumer
 releases them.
 
 For the detailed IRQ spread and allocation scheme  please see last patch:
 ("net/mlx5: Round-Robin EQs over IRQs")
 -----BEGIN PGP SIGNATURE-----
 
 iQEzBAABCAAdFiEEGhZs6bAKwk/OTgTpSD+KveBX+j4FAmDIJUkACgkQSD+KveBX
 +j7tgQf+KtxzniuEY+JgbGWWyQvglx88S6WfhTOhZZllm2QXa2wWX24mz/AdYc0x
 QCT6yUzvaeaHPNpw/KwCw1IKpB9dlT+wIBD9NCEqtHqj+bVz+ioL/OlM5VJj+wC2
 kp+EjYsQbwgZIM40JgLLu2uzLy/5w7a1v9Rj0l4mLRZqPmrqeKrIAsVkVutaxtPg
 PtECBag4XtYERMXOfKohnXanwjW6ZyYQ0Yal76jNqoXXgy5dHr/JJDZQZTDURt7S
 3ex0gwTZwHfOLFQdRzD+U0kuC2/6sHMfeVrKO6QxuG/gihYe8FXEQ4qVSJmgXANP
 VH6n1Vk5IhaMzYKfGFb2OGOWanAVIA==
 =z0x7
 -----END PGP SIGNATURE-----

Merge tag 'mlx5-updates-2021-06-14' of git://git.kernel.org/pub/scm/linux/kernel/git/saeed/linux

Saeed Mahameed says:

====================
mlx5-updates-2021-06-14

1) Trivial Lag refactroing in preparation for upcomming Single FDB lag feature
 - First 3 patches

2) Scalable IRQ distriburion for Sub-functions

A subfunction (SF) is a lightweight function that has a parent PCI
function (PF) on which it is deployed.

Currently, mlx5 subfunction is sharing the IRQs (MSI-X) with their
parent PCI function.

Before this series the PF allocates enough IRQs to cover
all the cores in a system, Newly created SFs will re-use all the IRQs
that the PF has allocated for itself.
Hence, the more SFs are created, there are more EQs per IRQs. Therefore,
whenever we handle an interrupt, we need to pull all SFs EQs and PF EQs
instead of PF EQs without SFs on the system. This leads to a hard impact
on the performance of SFs and PF.

For example, on machine with:
Intel(R) Xeon(R) CPU E5-2697 v3 @ 2.60GHz with 56 cores.
PCI Express 3 with BW of 126 Gb/s.
ConnectX-5 Ex; EDR IB (100Gb/s) and 100GbE; dual-port QSFP28; PCIe4.0 x16.

test case: iperf TX BW single CPU, affinity of app and IRQ are the same.
PF only: no SFs on the system, 56 IRQs.
SF (before), 250 SFs Sharing the same 56 IRQs .
SF (now),    250 SFs + 255 avaiable IRQs for the NIC. (please see IRQ spread scheme below).

	    application SF-IRQ  channel   BW(Gb/sec)         interrupts/sec
            iperf TX            affinity
PF only     cpu={0}     cpu={0} cpu={0}   79                 8200
SF (before) cpu={0}     cpu={0} cpu={0}   51.3 (-35%)        9500
SF (now)    cpu={0}     cpu={0} cpu={0}   78 (-2%)           8200

command:
$ taskset -c 0 iperf -c 11.1.1.1 -P 3 -i 6 -t 30 | grep SUM

The different between the SF examples is that before this series we
allocate num_cpus (56) IRQs, and all of them were shared among the PF
and the SFs. And after this series, we allocate 255 IRQs, and we spread
the SFs among the above IRQs. This have significantly decreased the load
on each IRQ and the number of EQs per IRQ is down by 95% (251->11).

In this patchset the solution proposed is to have a dedicated IRQ pool
for SFs to use. the pool will allocate a large number of IRQs
for SFs to grab from in order to minimize irq sharing between the
different SFs.
IRQs will not be requested from the OS until they are 1st requested by
an SF consumer, and will be eventually released when the last SF consumer
releases them.

For the detailed IRQ spread and allocation scheme  please see last patch:
("net/mlx5: Round-Robin EQs over IRQs")
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
This commit is contained in:
David S. Miller 2021-06-15 11:14:21 -07:00
commit f0c227c7df
17 changed files with 794 additions and 435 deletions
Download patch file Download diff file Expand all files Collapse all files