lkml.org 
[lkml]   [2024]   [Aug]   [13]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [xdp-hints] Re: [PATCH RFC bpf-next 32/52] bpf, cpumap: switch to GRO from netif_receive_skb_list()
From


On 13/08/2024 03.33, Jakub Kicinski wrote:
> On Fri, 9 Aug 2024 14:20:25 +0200 Alexander Lobakin wrote:
>> But I think one solution could be:
>>
>> 1. We create some generic structure for cpumap, like
>>
>> struct cpumap_meta {
>> u32 magic;
>> u32 hash;
>> }
>>
>> 2. We add such check in the cpumap code
>>
>> if (xdpf->metalen == sizeof(struct cpumap_meta) &&
>> <here we check magic>)
>> skb->hash = meta->hash;
>>
>> 3. In XDP prog, you call Rx hints kfuncs when they're available, obtain
>> RSS hash and then put it in the struct cpumap_meta as XDP frame metadata.
>
> I wonder what the overhead of skb metadata allocation is in practice.
> With Eric's "return skb to the CPU of origin" we can feed the lockless
> skb cache one the right CPU, and also feed the lockless page pool
> cache. I wonder if batched RFS wouldn't be faster than the XDP thing
> that requires all the groundwork.

I explicitly developed CPUMAP because I was benchmarking Receive Flow
Steering (RFS) and Receive Packet Steering (RPS), which I observed was
the bottleneck. The overhead was too large on the RX-CPU and bottleneck
due to RFS and RPS maintaining data structures to avoid Out-of-Order
packets. The Flow Dissector step was also a limiting factor.

By bottleneck I mean it didn't scale, as RX-CPU packet per second
processing speeds was too low compared to the remote-CPU pps.
Digging in my old notes, I can see that RPS was limited to around 4.8
Mpps (and I have a weird disabling part of it showing 7.5Mpps). In [1]
remote-CPU could process (starts at) 2.7 Mpps when dropping UDP packet
due to UdpNoPorts configured (and baseline 3.3 Mpps if not remote), thus
it only scales up-to 1.78 remote-CPUs. [1] shows how optimizations
brings remote-CPU to handle 3.2Mpps (close non-remote to 3.3Mpps
baseline). In [2] those optimizations bring remote-CPU to 4Mpps (for
UdpNoPorts case). XDP RX-redirect in [1]+[2] was around 19Mpps (which
might be lower today due to perf paper cuts).

[1]
https://github.com/xdp-project/xdp-project/blob/master/areas/cpumap/cpumap02-optimizations.org
[2]
https://github.com/xdp-project/xdp-project/blob/master/areas/cpumap/cpumap03-optimizations.org

The benefits Eric's "return skb to the CPU of origin" should help
improve the case for the remote-CPU, as I was seeing some bottlenecks in
how we returned the memory.

--Jesper

\
 
 \ /
  Last update: 2024-08-13 11:53    [from the cache]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and my Meterkast|Read the blog