Message-ID: <fe42bbef-047c-49da-b9ff-a7806820ae0e@uwaterloo.ca>
Date: Mon, 25 Aug 2025 15:45:05 -0400
From: Martin Karsten <mkarsten@...terloo.ca>
To: Samiullah Khawaja <skhawaja@...gle.com>
Cc: Jakub Kicinski <kuba@...nel.org>, "David S . Miller"
 <davem@...emloft.net>, Eric Dumazet <edumazet@...gle.com>,
 Paolo Abeni <pabeni@...hat.com>, almasrymina@...gle.com, willemb@...gle.com,
 Joe Damato <joe@...a.to>, netdev@...r.kernel.org
Subject: Re: [PATCH net-next v7 0/2] Add support to do threaded napi busy poll

On 2025-08-25 14:53, Samiullah Khawaja wrote:
> On Mon, Aug 25, 2025 at 10:41 AM Martin Karsten <mkarsten@...terloo.ca> wrote:
>>
>> On 2025-08-25 13:20, Samiullah Khawaja wrote:
>>> On Sun, Aug 24, 2025 at 5:03 PM Martin Karsten <mkarsten@...terloo.ca> wrote:
>>>>
>>>> On 2025-08-24 17:54, Samiullah Khawaja wrote:
>>>>> Extend the already existing support of threaded napi poll to do continuous
>>>>> busy polling.
>>>>>
>>>>> This is used for doing continuous polling of napi to fetch descriptors
>>>>> from backing RX/TX queues for low latency applications. Allow enabling
>>>>> of threaded busypoll using netlink so this can be enabled on a set of
>>>>> dedicated napis for low latency applications.
>>>>>
>>>>> Once enabled user can fetch the PID of the kthread doing NAPI polling
>>>>> and set affinity, priority and scheduler for it depending on the
>>>>> low-latency requirements.
>>>>>
>>>>> Currently threaded napi is only enabled at device level using sysfs. Add
>>>>> support to enable/disable threaded mode for a napi individually. This
>>>>> can be done using the netlink interface. Extend `napi-set` op in netlink
>>>>> spec that allows setting the `threaded` attribute of a napi.
>>>>>
>>>>> Extend the threaded attribute in napi struct to add an option to enable
>>>>> continuous busy polling. Extend the netlink and sysfs interface to allow
>>>>> enabling/disabling threaded busypolling at device or individual napi
>>>>> level.
>>>>>
>>>>> We use this for our AF_XDP based hard low-latency usecase with usecs
>>>>> level latency requirement. For our usecase we want low jitter and stable
>>>>> latency at P99.
>>>>>
>>>>> Following is an analysis and comparison of available (and compatible)
>>>>> busy poll interfaces for a low latency usecase with stable P99. Please
>>>>> note that the throughput and cpu efficiency is a non-goal.
>>>>>
>>>>> For analysis we use an AF_XDP based benchmarking tool `xdp_rr`. The
>>>>> description of the tool and how it tries to simulate the real workload
>>>>> is following,
>>>>>
>>>>> - It sends UDP packets between 2 machines.
>>>>> - The client machine sends packets at a fixed frequency. To maintain the
>>>>>      frequency of the packet being sent, we use open-loop sampling. That is
>>>>>      the packets are sent in a separate thread.
>>>>> - The server replies to the packet inline by reading the pkt from the
>>>>>      recv ring and replies using the tx ring.
>>>>> - To simulate the application processing time, we use a configurable
>>>>>      delay in usecs on the client side after a reply is received from the
>>>>>      server.
>>>>>
>>>>> The xdp_rr tool is posted separately as an RFC for tools/testing/selftest.
>>>>>
>>>>> We use this tool with following napi polling configurations,
>>>>>
>>>>> - Interrupts only
>>>>> - SO_BUSYPOLL (inline in the same thread where the client receives the
>>>>>      packet).
>>>>> - SO_BUSYPOLL (separate thread and separate core)
>>> This one uses separate thread and core for polling the napi.
>>
>> That's not what I am referring to below.
>>
>> [snip]
>>
>>>>> | Experiment | interrupts | SO_BUSYPOLL | SO_BUSYPOLL(separate) | NAPI threaded |
>>>>> |---|---|---|---|---|
>>>>> | 12 Kpkt/s + 0us delay | | | | |
>>>>> |  | p5: 12700 | p5: 12900 | p5: 13300 | p5: 12800 |
>>>>> |  | p50: 13100 | p50: 13600 | p50: 14100 | p50: 13000 |
>>>>> |  | p95: 13200 | p95: 13800 | p95: 14400 | p95: 13000 |
>>>>> |  | p99: 13200 | p99: 13800 | p99: 14400 | p99: 13000 |
>>>>> | 32 Kpkt/s + 30us delay | | | | |
>>>>> |  | p5: 19900 | p5: 16600 | p5: 13100 | p5: 12800 |
>>>>> |  | p50: 21100 | p50: 17000 | p50: 13700 | p50: 13000 |
>>>>> |  | p95: 21200 | p95: 17100 | p95: 14000 | p95: 13000 |
>>>>> |  | p99: 21200 | p99: 17100 | p99: 14000 | p99: 13000 |
>>>>> | 125 Kpkt/s + 6us delay | | | | |
>>>>> |  | p5: 14600 | p5: 17100 | p5: 13300 | p5: 12900 |
>>>>> |  | p50: 15400 | p50: 17400 | p50: 13800 | p50: 13100 |
>>>>> |  | p95: 15600 | p95: 17600 | p95: 14000 | p95: 13100 |
>>>>> |  | p99: 15600 | p99: 17600 | p99: 14000 | p99: 13100 |
>>>>> | 12 Kpkt/s + 78us delay | | | | |
>>>>> |  | p5: 14100 | p5: 16700 | p5: 13200 | p5: 12600 |
>>>>> |  | p50: 14300 | p50: 17100 | p50: 13900 | p50: 12800 |
>>>>> |  | p95: 14300 | p95: 17200 | p95: 14200 | p95: 12800 |
>>>>> |  | p99: 14300 | p99: 17200 | p99: 14200 | p99: 12800 |
>>>>> | 25 Kpkt/s + 38us delay | | | | |
>>>>> |  | p5: 19900 | p5: 16600 | p5: 13000 | p5: 12700 |
>>>>> |  | p50: 21000 | p50: 17100 | p50: 13800 | p50: 12900 |
>>>>> |  | p95: 21100 | p95: 17100 | p95: 14100 | p95: 12900 |
>>>>> |  | p99: 21100 | p99: 17100 | p99: 14100 | p99: 12900 |
>>>>>
>>>>>     ## Observations
>>>>
>>>> Hi Samiullah,
>>>>
>>> Thanks for the review
>>>> I believe you are comparing apples and oranges with these experiments.
>>>> Because threaded busy poll uses two cores at each end (at 100%), you
>>> The SO_BUSYPOLL(separate) column is actually running in a separate
>>> thread and using two cores. So this is actually comparing apples to
>>> apples.
>>
>> I am not referring to SO_BUSYPOLL, but to the column labelled
>> "interrupts". This is single-core, yes?
>>
>>>> should compare with 2 pairs of xsk_rr processes using interrupt mode,
>>>> but each running at half the rate. I am quite certain you would then see
>>>> the same latency as in the baseline experiment - at much reduced cpu
>>>> utilization.
>>>>
>>>> Threaded busy poll reduces p99 latency by just 100 nsec, while
>>> The table in the experiments show much larger differences in latency.
>>
>> Yes, because all but the first experiment add processing delay to
>> simulate 100% load and thus most likely show queuing effects.
>>
>> Since "interrupts" uses just one core and "NAPI threaded" uses two, a
>> fair comparison would be for "interrupts" to run two pairs of xsk_rr at
>> half the rate each. Then the load would be well below 100%, no queueing,
>> and latency would probably go back to the values measured in the "0us
>> delay" experiments. At least that's what I would expect.
> Two set of xsk_rr will go to two different NIC queues with two
> different interrupts (I think). That would be comparing apples to
> oranges, as all the other columns use a single NIC queue. Having
> (Forcing user to have) two xsk sockets to deliver packets at a certain
> rate is a completely different use case.

I don't think a NIC queue is a more critical resource than a CPU core?

And the rest depends on the actual application that would be using the 
service. The restriction to xsk_rr and its particulars is because that's 
the benchmark you provided.
>> Reproduction is getting a bit difficult, because you haven't updated the
>> xsk_rr RFC and judging from the compilation error, maybe not built/run
>> these experiments for a while now? It would be nice to have a working
>> reproducible setup.
> Oh. Let me check the xsk_rr and see whether it is outdated. I will
> send out another RFC for it if it's outdated.
>>
>>>> busy-spinning two cores, at each end - not more not less. I continue to
>>>> believe that this trade-off and these limited benefits need to be
>>>> clearly and explicitly spelled out in the cover letter.
>>> Yes, if you just look at the first row of the table then there is
>>> virtually no difference.
>> I'm not sure what you mean by this. I compare "interrupts" with "NAPI
>> threaded" for the case "12 Kpkt/s + 0us delay" and I have explained why
>> I believe the other experiments are not meaningful.
> Yes that is exactly what I am disagreeing with. I don't think other
> rows are "not meaningful". The xsk_rr is trying to "simulate the
> application processing" by adding a cpu delay and the table clearly
> shows the comparison between various mechanisms and how they perform
> with in load.

But these experiments only look at cases with almost exactly 100% load. 
As I mentioned in a previous round, this is highly unlikely for a 
latency-critical service and thus it seems contrived. Once you go to 
100% load and see queueing effects, you also need to look left and right 
to investigate other load and system settings.

Maybe this means the xsk_rr tool is not a good enough benchmark?

Thanks,
Martin

Powered by blists - more mailing lists