Skip site navigation (1)Skip section navigation (2)
Date:      Mon, 12 Nov 2018 12:51:29 +0000
From:      "Bjoern A. Zeeb" <bzeeb-lists@lists.zabbadoz.net>
To:        "Eugene Grosbein" <eugen@grosbein.net>
Cc:        "Wolfgang Zenker" <wolfgang@lyxys.ka.sub.org>, freebsd-net@freebsd.org
Subject:   Re: epair(4) and net.route.netisr_maxqlen
Message-ID:  <C5345D14-DD58-4014-A625-082492FE2359@lists.zabbadoz.net>
In-Reply-To: <5afe5838-d2df-009d-af85-99e793f1927c@grosbein.net>
References:  <20181111232320.GA87730@lyxys.ka.sub.org> <5afe5838-d2df-009d-af85-99e793f1927c@grosbein.net>

next in thread | previous in thread | raw e-mail | index | archive | help
On 12 Nov 2018, at 3:44, Eugene Grosbein wrote:

My additions are mostly for Wolfgang,

> 12.11.2018 6:23, Wolfgang Zenker wrote:
>
>> on a jail with quite a lot of somewhat bursty network traffic I get
>> warnings from netdata recently about packets being dropped because of
>> net.route.netisr_maxqlen being to small. Before I start setting this
>> value to some random value I'ld like to find out what it actually 
>> means.
>> A search for documentation turned up nothing; a look at the sources
>> found that it is used for the size of a "software interrupt queue" in
>> epair(4). But what does it mean? And does this give me enough
>> information to find a good value to set for this sysctl?
>
> netisr packet queues keep packets received by the interface and
> not yet processed by destined subsystem or userland application
> that may be short of CPU cycles or blocked for some reason.
>
> First, the system won't allow you to raise net.route.netisr_maxqlen 
> over the limit net.isr.maxqlimit.
> The limit itself can be changed with /boot/loader.conf and reboot.
> Default value of limit is 10240. I generally raise the limit upto 
> 102400
> for hosts with heavy/bursty traffic. Note that actual increase of 
> net.route.netisr_maxqlen
> somewhat increases usage of kernel memory and that could be important 
> for 32 bit kernel
> and/or system with very low amount of RAM.
>
> There may be several netisr packet queues in the system and raising 
> net.route.netisr_maxqlen
> allows all of them to grow. epair(4) has its own setting 
> net.link.epair.netisr_maxqlen
> that defaults to net.route.netisr_maxqlen if unset, so you may be 
> start experimenting with
> net.link.epair.netisr_maxqlen first, instead of system global 
> net.route.netisr_maxqlen.
>
> Don't set net.route.netisr_maxqlen to random value but double its 
> current value
> and see if that would be enough. If not, double it again. If 4096 
> apears not enough,
> you should check your applications why they can't keep with incoming 
> traffic rate.

Also if you have multiple epair interface-combinations and a multi-core 
CPU you might also want to try (which never became default I think) to 
experiment with these settings in loader.conf:

net.isr.bindthreads=1
net.isr.maxthreads=256         # net.isr will always reduce it to 
mp_cpus

which should help balancing the load across the cores.  Note: these 
changes also affect all other possible traffic going through the netisr 
framework.

The netisr(9) man page has some documentation of these fields but not 
everything.  The source code does have a lot of comments and if someone 
would improve the man page that might be a good start:
https://svnweb.freebsd.org/base/head/sys/net/netisr.c?annotate=326272#l159


netstat -Q is also a good source for monitoring and diagnostics.


/bz



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?C5345D14-DD58-4014-A625-082492FE2359>