C
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
src
tests
.gitignore
DESIGN
Makefile.am
README
TODO
autogen.sh
configure.ac
irqd.8
irqd.spec
irqdrc.5
rc.irqd

README

README
------

This is irqd, an alternative IRQ balancer for Linux kernels.  It
balances the IRQs more or less evenly across the CPUs you have in your
system.

As of version 0.6.2 it requires glib-2.0 and libnl-3.2 to compile.

It currently relies on the Receive Packet Steering (RPS)
implementation of newer Linux kernels, it is of less use currently if
RPS is not available.  For details on RPS see the very good article
from LWN: http://lwn.net/Articles/362339.


STRATEGY
--------

It currently implements the most gainfull cases only:

1. each multiqueue NIC is balanced over as much CPUs as possible.

2. for singlequeue NICs start with assigning two CPUs per IRQs.  This
   way you should experience an improvement over the existing
   right from the start.

3. for singlequeue NICs it will currently assign up to four CPUs
   per IRQ.

4. for non-RPS kernels it just balances multiqueue NICs.  Singlequeue
   NICs on non-RPS kernels will be addressed in the future.

Implementation of irqd is currently ongoing.


RUNNING
-------

irqd is currently very minimalistic, only serving a very dedicated
purpose.  It reports the current IRQ mapping to the different
CPUs to the file /var/lib/misc/irqd.cpumap.  Example contents:

cpu0: eth0:0 eth1:2                                                           
cpu1: eth0:0 eth1:3                                                           
cpu2: eth1:0                                                                  
cpu3: eth1:1                                                                  

Here eth0 being a singlequeue NIC and eth1 being multiqueue (with four
queues).


BUGS
----

irqd may contain bugs.  Any bug report and/or feature request is
welcomed, but please look at the TODO file contained before
making a request.

-- 
Holger Eitzenberger <holger@eitzenberger.org>