Optimizing kernel compilation / alignments for network performance

[Rafał Miłecki]

On 6.05.2022 10:45, Arnd Bergmann wrote:
> On Fri, May 6, 2022 at 9:44 AM Rafał Miłecki <zajec5 at gmail.com> wrote:
>>
>> On 5.05.2022 18:04, Andrew Lunn wrote:
>>>> you'll see that most used functions are:
>>>> v7_dma_inv_range
>>>> __irqentry_text_end
>>>> l2c210_inv_range
>>>> v7_dma_clean_range
>>>> bcma_host_soc_read32
>>>> __netif_receive_skb_core
>>>> arch_cpu_idle
>>>> l2c210_clean_range
>>>> fib_table_lookup
>>>
>>> There is a lot of cache management functions here.
> 
> Indeed, so optimizing the coherency management (see Felix' reply)
> is likely to help most in making the driver faster, but that does not
> explain why the alignment of the object code has such a big impact
> on performance.
> 
> To investigate the alignment further, what I was actually looking for
> is a comparison of the profile of the slow and fast case. Here I would
> expect that the slow case spends more time in one of the functions
> that don't deal with cache management (maybe fib_table_lookup or
> __netif_receive_skb_core).
> 
> A few other thoughts:
> 
> - bcma_host_soc_read32() is a fundamentally slow operation, maybe
>    some of the calls can turned into a relaxed read, like the readback
>    in bgmac_chip_intrs_off() or the 'poll again' at the end bgmac_poll(),
>    though obviously not the one in bgmac_dma_rx_read().
>    It may be possible to even avoid some of the reads entirely, checking
>    for more data in bgmac_poll() may actually be counterproductive
>    depending on the workload.
> 
> - The higher-end networking SoCs are usually cache-coherent and
>    can avoid the cache management entirely. There is a slim chance
>    that this chip is designed that way and it just needs to be enabled
>    properly. Most low-end chips don't implement the coherent
>    interconnect though, and I suppose you have checked this already.
> 
> - bgmac_dma_rx_update_index() and bgmac_dma_tx_add() appear
>    to have an extraneous dma_wmb(), which should be implied by the
>    non-relaxed writel() in bgmac_write().
> 
> - accesses to the DMA descriptor don't show up in the profile here,
>    but look like they can get misoptimized by the compiler. I would
>    generally use READ_ONCE() and WRITE_ONCE() for these to
>    ensure that you don't end up with extra or out-of-order accesses.
>    This also makes it clearer to the reader that something special
>    happens here.
> 
>>> Might sound odd,
>>> but have you tried disabling SMP? These cache functions need to
>>> operate across all CPUs, and the communication between CPUs can slow
>>> them down. If there is only one CPU, these cache functions get simpler
>>> and faster.
>>>
>>> It just depends on your workload. If you have 1 CPU loaded to 100% and
>>> the other 3 idle, you might see an improvement. If you actually need
>>> more than one CPU, it will probably be worse.
>>
>> It seems to lower my NAT speed from ~362 Mb/s to 320 Mb/s but it feels
>> more stable now (lower variations). Let me spend some time on more
>> testing.
>>
>>
>> FWIW during all my tests I was using:
>> echo 2 > /sys/class/net/eth0/queues/rx-0/rps_cpus
>> that is what I need to get similar speeds across iperf sessions
>>
>> With
>> echo 0 > /sys/class/net/eth0/queues/rx-0/rps_cpus
>> my NAT speeds were jumping between 4 speeds:
>> 273 Mbps / 315 Mbps / 353 Mbps / 425 Mbps
>> (every time I started iperf kernel jumped into one state and kept the
>>    same iperf speed until stopping it and starting another session)
>>
>> With
>> echo 1 > /sys/class/net/eth0/queues/rx-0/rps_cpus
>> my NAT speeds were jumping between 2 speeds:
>> 284 Mbps / 408 Mbps
> 
> Can you try using 'numactl -C' to pin the iperf processes to
> a particular CPU core? This may be related to the locality of
> the user process relative to where the interrupts end up.

I run iperf on x86 machines connected to router's WAN and LAN ports.
It's meant to emulate end user just downloading from / uploading to
Internet some data.

Router's only task is doing masquarade NAT here.