|
Hi!
> I'm thinking I will be using the keepalived implementation and use the
> direct routing (LVS-DR) method for forwarding (just for some
> background...)
Exactly what we do. We use the same GE interface for both
incoming and outgoing connections these days (we used to have
dedicated interfaces, but that wasn't actually necessary).
> I want to pick the hardware with this in mind; will LVS benefit most
> from more RAM? more CPU? what things should I consider (I'm a newbie
> at this type of service, so I'd like to ask the people using it what
> seems to be the best approach)
In my experience, IPVS benefits most from interrupt rate.
I've yet to run out of RAM on any of my load balancers, but I've
seen dents in conn rate curves when getting near the maximum
interrupt rate on older machines (or lower quality
chipsets/mainboards).
As a data point, my hardest-working server reports (in
ip_vs_stat) that it has 5041 Conns/s and 32147 Pkts/s (incoming).
Somehow those number seem a bit high to me, but then, a
ipvsadm reports 20k P/s and 3295 C/s on the main farm on that
balancer. Traffic is 3266361 B/s.
As for hardware, we use dual-cpu PIVs @ 2.8GHz with 1G of RAM.
NICs are two Broadcom NetXtreme (BCM5703X) Gig Ethernet cards on
copper GE. All of them work allong nicely, I've yet to bounce my
head into their performance roof.
The biggest farm on the busiest balancer runs at about 3000-3500
connections per second or about 7.5G Conns per month.
> I'd be looking for the "ideal" situation, but if it could be ordered
> based on the pressure points, that'd help a lot. i.e. if processor
> cache size really matters, clock speed, memory speed, amount,
> configuration, etc.
Try to go for GE cards, those usually are capable of interrupt
coalescing, which helps drop the interrupt rate. As I said that's
the only metric I've yet run into limits with.
> Right now, over web alone, it appears I'm doing 12-15 million requests
> per day. I expect this number to grow a lot more. I currently have 6
> webservers serving up content behind my existing [blackbox/ISP
> provided] load balancing service. I want plenty of headroom to grow.
The farm I mentioned earlier is at about 250-300 M conns per day
- and it's not the only one on that particular balancer.
> I do intend on building 2 identical machines so I have a slave
> director for redundancy, as well.
Good idea, we do the same.
> Thanks in advance.
Hope I helped.
Regards,
Tobias
--
export DISPLAY=vt100
|
