> -----Original Message----- > From: Andrew Lunn <andrew@xxxxxxx> > Sent: Thursday, March 21, 2024 12:42 PM > To: Selvamani Rajagopal <Selvamani.Rajagopal@xxxxxxxxxx> > Cc: Parthiban.Veerasooran@xxxxxxxxxxxxx; davem@xxxxxxxxxxxxx; > edumazet@xxxxxxxxxx; kuba@xxxxxxxxxx; pabeni@xxxxxxxxxx; > horms@xxxxxxxxxx; saeedm@xxxxxxxxxx; anthony.l.nguyen@xxxxxxxxx; > netdev@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx; corbet@xxxxxxx; > linux-doc@xxxxxxxxxxxxxxx; robh+dt@xxxxxxxxxx; > krzysztof.kozlowski+dt@xxxxxxxxxx; conor+dt@xxxxxxxxxx; > devicetree@xxxxxxxxxxxxxxx; Horatiu.Vultur@xxxxxxxxxxxxx; > ruanjinjie@xxxxxxxxxx; Steen.Hegelund@xxxxxxxxxxxxx; > vladimir.oltean@xxxxxxx; UNGLinuxDriver@xxxxxxxxxxxxx; > Thorsten.Kummermehr@xxxxxxxxxxxxx; Piergiorgio Beruto > <Pier.Beruto@xxxxxxxxxx>; Nicolas.Ferre@xxxxxxxxxxxxx; > benjamin.bigler@xxxxxxxxxxxxxxxxxxxxx > Subject: Re: [PATCH net-next v3 08/12] net: ethernet: oa_tc6: > implement transmit path to transfer tx ethernet frames > > [External Email]: This email arrived from an external source - Please > exercise caution when opening any attachments or clicking on links. > > > > > This second part is clearly an optimisation. If you have lots of > > > > full MTU packets, 1514 bytes, they take around 24 chunks. Having > > > > the last chunk only 1/2 full does not waste too much bandwidth. > > > > But if you are carrying lots of small packets, say voice, 130 > > > > bytes, the wasted bandwidth starts to add up. But is there a use > > > > case for 10Mbps of small packets? I doubt it. > > > Yes, for sure there is a possibility to get into this scenario and > > > the protocol also supports that. But as proposed by you below, let's > > > implement it as part of optimization later. > > > > > > > > So if you don't have the ability to combine two packets into one > > > > chunk, i would do that later. Lets get the basics merged first, it > > > > can be optimised later. > > > Yes, I agree with this proposal to get the basic version merged first. > > > > While latency is important, so is using the available bandwidth > > efficiently. Here is a suggestion. We know that the tx credit > > available basically tells us, how many chunks could be transmitted > without overflow. Instead of stopping the netif queue based on number > of skbs queued, why not stop the queue based on number of bytes > accumulated? Basically, at any given point of time, we enqueue the > tx_skb_q until we are have enough bytes to cross the threshold of (tc6- > >tc_credit * OA_TC6_CHUNK_PAYLOAD_SIZE). > > This way, during the next transmit, we could utilize the whole available > credits. Bandwidth utilization between bigger frames and smaller > frames would be not be vastly different. > > Please configure your email client to wrap emails at around 70 > characters. > > tc_credit is 5 bits. So it is a maximum of 32. > > A 1514 frame takes around 24 chunks. So you only need two full size > frames to consume all your possible credit. > > If you happen to have smaller voice packets, say 130 bytes, you need > three chunks to send it. So you might want to have 10 such packets on > hand in order to make use of all your credit. But if you have 10 voice > packets to send in a burst, your voice quality is going to be bad, they > should be 10ms to 20ms apart, not in a burst... > > I don't like the original idea of having lots of packets in a transmit queue. > But having 1/2 dozen should not be an issue. > > In general, we prefer things to be simple. We can then optimise later, > and use benchmarks to show the optimisations really do bring a benefit > to justify the added complexity. True. I should get some performance numbers to see where we are with the current code. That would be time to look at the improvement. > > Andrew