| 1 |
Hi, |
| 2 |
|
| 3 |
On Wed, 20 Jun 2007 11:15:04 -0300 |
| 4 |
"Daniel van Ham Colchete" <daniel.colchete@×××××.com> wrote: |
| 5 |
|
| 6 |
> I have a collocation server with a 2 Mbps bandwidth for my e-mail. I would |
| 7 |
> like to use SFQ to do a more fair split of this bandwidth between my users |
| 8 |
> (~500). I think SFQ is a very good idea: it will split the available |
| 9 |
> bandwidth between the current connections (almost) equally, and it's CPU |
| 10 |
> cheap. |
| 11 |
|
| 12 |
Yep. But remember: Connections, not users. |
| 13 |
|
| 14 |
> The problem is: from what I could read, SFQ can only read the available |
| 15 |
> bandwidth from the physical layer. Is it possible to tell SFQ something like |
| 16 |
> "I only have 2 Mbps, and not 100 Mbps, from eth0"? |
| 17 |
|
| 18 |
I didn't read anywhere it needs a bandwith parameter at all. And I |
| 19 |
think it really doesn't. From what I read in the |
| 20 |
source, /usr/src/linux/net/sched/sch_sfq.c, it merely just distributes |
| 21 |
incoming packets equally and it really only gets in between if the |
| 22 |
enqueuing rate exceeds the dequeuing rate -- by means of absolute |
| 23 |
packet buffer numbers, that is. |
| 24 |
|
| 25 |
In that case, using a TBF before giving packets to the SFQ would |
| 26 |
probably serve well. Also, you might want to use SFQ only for some |
| 27 |
connections, e.g. give precedence to DNS queries or similar stuff. |
| 28 |
|
| 29 |
Just put a TBF as root discipline, like (untested): |
| 30 |
$ tc qdisc add dev eth0 root handle :1 tbf rate 1920kbit latency 50ms burst 4096 |
| 31 |
and then set up a SFQ discipline: |
| 32 |
$ tc qdisc add dev eth0 parent 1:1 sfq |
| 33 |
|
| 34 |
If OTOH you decide to use PRIO scheduling for interactive traffic, you |
| 35 |
can probably fully replace the TBF by the HTB discipline and use SFQ at |
| 36 |
the ends of the subclasses. |
| 37 |
|
| 38 |
-hwh |
| 39 |
-- |
| 40 |
gentoo-user@g.o mailing list |
Archives