[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <87pln99a28.fsf@toke.dk>
Date: Tue, 05 Nov 2024 12:47:27 +0100
From: Toke Høiland-Jørgensen <toke@...hat.com>
To: Simon Horman <horms@...nel.org>
Cc: Qingfang Deng <dqfext@...il.com>, "David S. Miller"
<davem@...emloft.net>, Eric Dumazet <edumazet@...gle.com>, Jakub Kicinski
<kuba@...nel.org>, Paolo Abeni <pabeni@...hat.com>,
netdev@...r.kernel.org, linux-kernel@...r.kernel.org,
linux-ppp@...r.kernel.org
Subject: Re: [RFC PATCH net-next] net: ppp: convert to IFF_NO_QUEUE
Simon Horman <horms@...nel.org> writes:
> + Toke
>
> On Tue, Oct 29, 2024 at 06:36:56PM +0800, Qingfang Deng wrote:
>> When testing the parallel TX performance of a single PPPoE interface
>> over a 2.5GbE link with multiple hardware queues, the throughput could
>> not exceed 1.9Gbps, even with low CPU usage.
>>
>> This issue arises because the PPP interface is registered with a single
>> queue and a tx_queue_len of 3. This default behavior dates back to Linux
>> 2.3.13, which was suitable for slower serial ports. However, in modern
>> devices with multiple processors and hardware queues, this configuration
>> can lead to congestion.
>>
>> For PPPoE/PPTP, the lower interface should handle qdisc, so we need to
>> set IFF_NO_QUEUE. For PPP over a serial port, we don't benefit from a
>> qdisc with such a short TX queue, so handling TX queueing in the driver
>> and setting IFF_NO_QUEUE is more effective.
>>
>> With this change, PPPoE interfaces can now fully saturate a 2.5GbE link.
>>
>> Signed-off-by: Qingfang Deng <dqfext@...il.com>
>
> Hi Toke,
>
> I'm wondering if you could offer an opinion on this.
Hi Simon
Thanks for bringing this to my attention; I'll reply to the parent
directly :)
-Toke
Powered by blists - more mailing lists