Re: [virtio-comment] Re: [PATCH RESEND] virtio-net: support setting coalescing params for multiple vqs

[Heng Qi <hengqi@linux.alibaba.com>]

å 2023/12/22 äå4:20, Michael S. Tsirkin åé:
> On Fri, Dec 22, 2023 at 02:45:21PM +0800, Heng Qi wrote:
> > > > For the current upstream code, the overhead is 256 kicks + 256*8 DMA times.
> > > > The overhead of batch cmds is 1 kick + 256*8 DMA times.
> > > This can furtherly optimized by using IN_ORDER. The DMA can be batched as well.
> > DMA batch reduces the number of DMA requests, but not the actual number of
> > DMAs.
> > For out DPU, how much to batch and when to end the batch also need to be
> > evaluated in detail.
> And what is the result of this evaluation?

Itâs not easy. Not having a good start and end time means that DMA batch 
will increase the response delay
of the entire machine. And DMA batch will not reduce the number of DMA 
times.

>   Again all this is slow path
> stuff. Speed only matters if the operation takes many milliseconds.

Yes, when there are many queues and the device responds slowly, the CPU 
where dim_work in
the driver is located may be kept waiting. This is the overhead 
introduced by DIM enablding.
We want to try our best to avoid the impact of DIM enabling. And, due to 
DPU resource limitations,
we want to better support multiple VMs.

Thanks.