Re: wheezy irqbalance
On 5/10/2013 3:39 AM, Guy Marcenac wrote:
> Le 09/05/2013 17:34, Stan Hoeppner a écrit :
>> Or simply cron a stop/start of irqbalance some x minutes after bootup.
>
> Thank you Stan
>
> I tryied restarting irqbalance on my running system with
> /etc/init.d/irqbalance restart
> But the warning still appears in the log...
>
> I think I will wait for an update since I understand that not having
> irqbalance running does not harm the system.
It was discovered some time back that IO would not scale as it should on
many/most busy systems, because all the interrupts were being routed by
the PC platform chipsets to CPU0 and swamping it, limiting IO throughput
to a fraction of the hardware capability. This default IRQ routing to
CPU0 is an Intel platform spec from long ago and every PC based system
does this. This is especially a problem on systems such as those
sporting lots of sockets and cores, such as an AMD system w/4 sockets/48
low TDP cores at only 1.7GHz, with say 10 PCIe x8/16 slots. It doesn't
take that much IO across the HBAs to saturate a 1.7GHz Magny Cours core
with IRQ processing.
irqbalance was written to remap the chipset routing table such that IRQs
from an HBA could be mapped to a specific CPU or a core within a CPU,
eliminating the congestion at CPU0.
Thus, unless a system has both multiple sockets/cores and multiple HBAs
(SAS, RAID, FC, Infiniband, GbE/10GbE), with lots of sustained IO and
thus interrupts, then irqbalance is really irrelevant. It's not needed
in desktops, nor most average servers, especially those with high clock
rate CPUs.
--
Stan
Reply to: