Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Mon, 5 Dec 2022 18:47:42 -0800
From:   Jakub Kicinski <kuba@kernel.org>
To:     <yang.yang29@zte.com.cn>
Cc:     <davem@davemloft.net>, <edumazet@google.com>, <pabeni@redhat.com>,
        <bigeasy@linutronix.de>, <imagedong@tencent.com>,
        <kuniyu@amazon.com>, <petrm@nvidia.com>, <liu3101@purdue.edu>,
        <wujianguo@chinatelecom.cn>, <netdev@vger.kernel.org>,
        <linux-kernel@vger.kernel.org>, <tedheadster@gmail.com>
Subject: Re: [PATCH linux-next] net: record times of netdev_budget exhausted
Message-ID: <20221205184742.0952fc75@kernel.org>
In-Reply-To: <202212061035074041030@zte.com.cn>
References: <20221205175314.0487527a@kernel.org>
        <202212061035074041030@zte.com.cn>
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Precedence: bulk

On Tue, 6 Dec 2022 10:35:07 +0800 (CST) yang.yang29@zte.com.cn wrote:
> The author of "Replace 2 jiffies with sysctl netdev_budget_usecs
> to enable softirq tuning" is Matthew Whitehead, he said this in
> git log: Constants used for tuning are generally a bad idea, especially
> as hardware changes over time...For example, a very fast machine
> might tune this to 1000 microseconds, while my regression testing
> 486DX-25 needs it to be 4000 microseconds on a nearly idle network
> to prevent time_squeeze from being incremented.

Let's just ignore that on the basis that it mentions prehistoric HW ;)

> And on my systems there are huge packets on the intranet, and we
> came accross with lots of time_squeeze. The idea is that, netdev_budget*
> are selections between throughput and real-time. If we care throughput
> and not care real-time so much, we may want bigger netdev_budget*.

But are you seeing actual performance wins in terms of throughput 
or latency? 

As I said time_squeeze is very noisy. In my experience it's very
sensitive to issues with jiffies, like someone masking interrupts on
the timekeeper CPU for a long time (which if you use cgroups happens
_a lot_ :/).

Have you tried threaded NAPI? (find files called 'threaded' in sysfs)
It will let you do any such tuning much more flexibly.

> In this scenario, we want to tune netdev_budget* and see their effect
> separately.
>
> By the way, if netdev_budget* are useless, should they be deleted?

Well, we can't be sure if there's really nobody that uses them :(
It's very risky to remove stuff that's exposed to user space.