Re: [RFC v2] net/core: add optional threading for rps backlog processing

From: Eric Dumazet
Date: Fri Feb 17 2023 - 09:39:13 EST


On Fri, Feb 17, 2023 at 2:40 PM Felix Fietkau <nbd@xxxxxxxx> wrote:
>
> On 17.02.23 13:57, Eric Dumazet wrote:
> > On Fri, Feb 17, 2023 at 1:35 PM Felix Fietkau <nbd@xxxxxxxx> wrote:
> >>
> >> On 17.02.23 13:23, Eric Dumazet wrote:
> >> > On Fri, Feb 17, 2023 at 11:06 AM Felix Fietkau <nbd@xxxxxxxx> wrote:
> >> >>
> >> >> When dealing with few flows or an imbalance on CPU utilization, static RPS
> >> >> CPU assignment can be too inflexible. Add support for enabling threaded NAPI
> >> >> for RPS backlog processing in order to allow the scheduler to better balance
> >> >> processing. This helps better spread the load across idle CPUs.
> >> >>
> >> >> Signed-off-by: Felix Fietkau <nbd@xxxxxxxx>
> >> >> ---
> >> >>
> >> >> RFC v2:
> >> >> - fix rebase error in rps locking
> >> >
> >> > Why only deal with RPS ?
> >> >
> >> > It seems you propose the sofnet_data backlog be processed by a thread,
> >> > instead than from softirq ?
> >> Right. I originally wanted to mainly improve RPS, but my patch does
> >> cover backlog in general. I will update the description in the next
> >> version. Does the approach in general make sense to you?
> >>
> >
> > I do not know, this seems to lack some (perf) numbers, and
> > descriptions of added max latencies and stuff like that :)
> I just ran some test where I used a MT7621 device (dual-core 800 MHz
> MIPS, 4 threads) as a router doing NAT without flow offloading.
>
> Using the flent RRUL test between 2 PCs connected through the router,
> I get these results:
>
> rps_threaded=0: (combined CPU idle time around 27%)
> avg median 99th % # data pts
> Ping (ms) ICMP : 26.08 28.70 54.74 ms 199
> Ping (ms) UDP BE : 1.96 24.12 37.28 ms 200
> Ping (ms) UDP BK : 1.88 15.86 27.30 ms 200
> Ping (ms) UDP EF : 1.98 31.77 54.10 ms 200
> Ping (ms) avg : 1.94 N/A N/A ms 200
> TCP download BE : 69.25 70.20 139.55 Mbits/s 200
> TCP download BK : 95.15 92.51 163.93 Mbits/s 200
> TCP download CS5 : 133.64 129.10 292.46 Mbits/s 200
> TCP download EF : 129.86 127.70 254.47 Mbits/s 200
> TCP download avg : 106.97 N/A N/A Mbits/s 200
> TCP download sum : 427.90 N/A N/A Mbits/s 200
> TCP totals : 864.43 N/A N/A Mbits/s 200
> TCP upload BE : 97.54 96.67 163.99 Mbits/s 200
> TCP upload BK : 139.76 143.88 190.37 Mbits/s 200
> TCP upload CS5 : 97.52 94.70 206.60 Mbits/s 200
> TCP upload EF : 101.71 106.72 147.88 Mbits/s 200
> TCP upload avg : 109.13 N/A N/A Mbits/s 200
> TCP upload sum : 436.53 N/A N/A Mbits/s 200
>
> rps_threaded=1: (combined CPU idle time around 16%)
> avg median 99th % # data pts
> Ping (ms) ICMP : 13.70 16.10 27.60 ms 199
> Ping (ms) UDP BE : 2.03 18.35 24.16 ms 200
> Ping (ms) UDP BK : 2.03 18.36 29.13 ms 200
> Ping (ms) UDP EF : 2.36 25.20 41.50 ms 200
> Ping (ms) avg : 2.14 N/A N/A ms 200
> TCP download BE : 118.69 120.94 160.12 Mbits/s 200
> TCP download BK : 134.67 137.81 177.14 Mbits/s 200
> TCP download CS5 : 126.15 127.81 174.84 Mbits/s 200
> TCP download EF : 78.36 79.41 143.31 Mbits/s 200
> TCP download avg : 114.47 N/A N/A Mbits/s 200
> TCP download sum : 457.87 N/A N/A Mbits/s 200
> TCP totals : 918.19 N/A N/A Mbits/s 200
> TCP upload BE : 112.20 111.55 164.38 Mbits/s 200
> TCP upload BK : 144.99 139.24 205.12 Mbits/s 200
> TCP upload CS5 : 93.09 95.50 132.39 Mbits/s 200
> TCP upload EF : 110.04 108.21 207.00 Mbits/s 200
> TCP upload avg : 115.08 N/A N/A Mbits/s 200
> TCP upload sum : 460.32 N/A N/A Mbits/s 200
>
> As you can see, both throughput and latency improve because load can be
> better distributed across CPU cores.
>

What happens if user threads are competing with your kthreads ?

It seems you are adding another variant of ksoftirqd.

More threads might look better in some cases, if we accept to be at
the mercy of process scheduling decisions.

NUMA affinities matter, I do not see how you are dealing with this.

Your patch assumes all cpus can participate in network processing,
but rps is fine grained:
Boxes with 128 or 256 cpus usually have different rps_mask per receive
queue, with 2 or 4 bits set in the per-rx-queue rps_mask.

Then, process_backlog() has been designed to run only from the cpu
tied to the per-cpu data (softnet_data)
There are multiple comments about this assumption, and various things
that would need to be changed
(eg sd_has_rps_ipi_waiting() would be wrong in its current implementation)