|
On Tue 03 May 2005 21:32:41 BST , Mark de Vries
<markdv.lvsuser@xxxxxxxxxx> wrote:
<snip>
I have 9 VIPs, each with the same 4 realservers so thats a lot of lines...
I'll leave some stuff out...
If I were you I would _seriously_ look at grouping these services using
fwmarks. It'll simplify your LVS tables by quite some margin - as you
may have seen from mine, which has 3 realservers providing DNS service
over TCP and UDP to... let me see... 46 VIPs. [thinks... "My list is
longer than yours" :p ]
I would say my rates are significantly higher... at peak hour the total is
even ~800 InPPS more...
Can't argue with that statement, your rates are most definitely higher
than mine. This has proved to be a "build and leave it alone" system in
LVS terms, so I don't very often look at the packet rates unless
something unusual happens.
Have you fiddled with the bucket size, BTW?
I'm going to compile a new kernel tomorrow and raise that one... Someone
else compiled this one and I see it's still at the default 4k. Hope that
helps. I think if it does it's still a bug though. Larger bucket size
should only make finding connections more efficient (less CPU cycles)...
Hrm... the reason I asked that was to see if it had been changed to
something silly, thus explaining why things were taking over in the way
they were.
Feel free to email offlist if you'd like to discuss specifics about the
cluster.
Graeme
|
