Dear List,
In order to address inaccurate results using either TBF or HTB according to:
http://www.docum.org/docum.org/faq/cache/40.html
or
"In order to increase the accuracy of the clock we can modify some
parameters of the kernel. The parameters we have tried to change
have been the PSCHED_CLOCK_SOURCE variable and the HZ variable. The observed
effect has been a small increase in the use
of the CPU, but a great improvement regarding the accuracy of the output bit
rate of the flows. The CBQ scheduler guarantees a
minimum bandwidth better and TBF obtains an output rate more adjusted to
what is required."
https://www.tlm.unavarra.es/~eduardo/publicaciones/20010403-alcom-english.pdf
It seems that the best setting for a dedicated box to shape traffic in the
network is to build a kernel with:
CONFIG_NET_SCHED=y
CONFIG_NET_SCH_CLK_CPU=y
But since NET_SCH_CLK_CPU depends on ((X86_TSC || X86_64) && !SMP),
Does this mean that the ideal Bandwidth Management box has to use only one
CPU?
The following link also suggests the use of preemption:
CONFIG_PREEMPT=y
CONFIG_PREEMPT_BKL=y
http://www.sigsegv.cx/qos-2.html
Any thoughts? What are the best kernel settings for a Bandwidth Management
box in bridge mode using X86 architecture?
Regards,
Mario Antonio
_______________________________________________
LARTC mailing list
LARTC@mailman.ds9a.nl
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc
|