Re: [PATCH] net :mana : Add per-cpu stats for MANA device

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Wed, 13 Mar 2024 19:57:20 -0700 Shradha Gupta wrote:
> Default interrupts affinity for each queue:
> 
>  25:          1        103          0    2989138  Hyper-V PCIe MSI 4138200989697-edge      mana_q0@pci:7870:00:00.0
>  26:          0          1    4005360          0  Hyper-V PCIe MSI 4138200989698-edge      mana_q1@pci:7870:00:00.0
>  27:          0          0          1    2997584  Hyper-V PCIe MSI 4138200989699-edge      mana_q2@pci:7870:00:00.0
>  28:    3565461          0          0          1  Hyper-V PCIe MSI 4138200989700-edge      mana_q3
> @pci:7870:00:00.0
> 
> As seen the CPU-queue mapping is not 1:1, Queue 0 and Queue 2 are both mapped 
> to cpu3. From this knowledge we can figure out the total RX stats processed by
> each CPU by adding the values of mana_q0 and mana_q2 stats for cpu3. But if
> this data changes dynamically using irqbalance or smp_affinity file edits, the
> above assumption fails. 
> 
> Interrupt affinity for mana_q2 changes and the affinity table looks as follows
>  25:          1        103          0    3038084  Hyper-V PCIe MSI 4138200989697-edge      mana_q0@pci:7870:00:00.0
>  26:          0          1    4012447          0  Hyper-V PCIe MSI 4138200989698-edge      mana_q1@pci:7870:00:00.0
>  27:     157181         10          1    3007990  Hyper-V PCIe MSI 4138200989699-edge      mana_q2@pci:7870:00:00.0
>  28:    3593858          0          0          1  Hyper-V PCIe MSI 4138200989700-edge      mana_q3@pci:7870:00:00.0 
> 
> And during this time we might end up calculating the per-CPU stats incorrectly,
> messing up the understanding of CPU usage by MANA driver that is consumed by 
> monitoring services. 

Like Stephen said, forget about irqbalance for networking.

Assume that the IRQs are affinitized and XPS set, correctly.

Now, presumably you can use your pcpu stats to "trade queues",
e.g. 4 CPUs / 4 queues, if CPU 0 insists on using queue 1
instead of queue 0, you can swap the 0 <> 1 assignment.
That's just an example of an "algorithm", maybe you have other
use cases. But if the problem is "user runs broken irqbalance"
the solution is not in the kernel...




[Index of Archives]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Photo]     [Yosemite News]     [Yosemite Photos]     [Linux Kernel]     [Linux SCSI]     [XFree86]

  Powered by Linux