Re: [PATCH 0/3] Basic deferred TX queue flushing infrastructure.

[Alexander Duyck <alexander.duyck@gmail.com>]

On 08/23/2014 01:28 PM, David Miller wrote:
> 
> Over time, and specifically and more recently at the Networking
> Workshop during Kernel SUmmit in Chicago, we have discussed the idea
> of having some way to optimize transmits of multiple TX packets at
> a time.
> 
> There are several areas of overhead that could be amortized with such
> schemes.  One has to do with locking and transactional overhead, the
> other has to do with device specific costs.
> 
> This patch set here is more aimed at device specific costs.
> 
> Typically a device queues up a packet in the TX queue and then has to
> do something to have the device start processing that new entry.
> Sometimes this is composed of doing an MMIO write to a "tail"
> register, and in other cases it can involve something as expensive as
> a hypervisor call.

The MMIO call isn't an issue until you encounter a locked operation, at
least on x86 architecture.  So this often shows up in perf traces as a
hit on the qdisc lock right after completing a transmit.  I've seen it
at around 20% of CPU utilization when I was doing routing work with ixgbe.

Thanks,

Alex