Congratulations! In sounds like you are beginning to get route-maps.
This is no small think. Every book that I've read so far says
"route-maps are beyond the scope of this book but here are some neat tricks"
Hmm. I'm going to have to lab both of these tomorrow to be 100% sure.
ICMP should be pretty straight forward, you should just need to add an
ACL to your route map to premit ICMP (all you are letting through is IP).
Cookie Switching looks interesting and it is on my to do list to look at
- I'll have to consult with my web guru. My guess is that the cookie is
getting used by the switch to tell it which which to go to. It looks
like cookie switching is trumping route maps on a first match basis.
You've probably already thought of this but what about forcing this data
to be ssl and then using session-id-switch for future connections? It is
a unique client/server keypair. If the customer you are worried about
comes from the big NAT/Proxy box, you can force their connections to be
written into SSL (mod_rewrite on the server)
Just a thought.
Ciao,
---Jack
Lazuardi Nasution wrote:
Hi Jack,
I have implemented route-maps. It is OK until I do cookie switching.
When I put cookie-switching, the second VIP never response the HTTP
request. The connection is OK but no response. If I remove the
cookie-switching, the traffic is normal again. I need the
cookie-switching for handling the clients behind big NAT/Proxy and
make sure that the request switched to the same server on the event of
link state changes.
Beside that, it seem that route-map cannot work for ICMP. I still
cannot Ping to second VIP.
Best regards,
On Mon, Dec 14, 2009 at 6:57 AM, Jack Stewart <[email protected]> wrote:
Hi Lazuardi,
Setting up a DNS response strategy seems like a good option - the idea is
sound and it is the foundation for GSLB for many vendors. I've read about
GSLB but never implemented it.
I applied route-maps to the management IP subnet for giggles and removed the
default route. It works but I suspect it isn't a best practice. A route-map
is more like pre-routing table. If you get a match in the route-map, then
the specific action is applied. Otherwise, it goes to the routing table.
Another way of thinking about it is that route-maps (or policy routing) have
the better metric but you still have your default and static routes.
Your setup isn't something that I've done but as long as it works, stick
with it. In my situation I have two class C subnets that are part of a
larger class B and default routes couldn't do the job.
Route maps are pretty cool and now that I understand them, I'm currently
exploring PBR as a mechanism for handling Email flows differently based on
source. Load balancing using DNS is way down the road for me.
Cheers,
---Jack
Lazuardi Nasution wrote:
Hi Jack,
Is this route-map can be applied to Mangement IPs ?
Currently, I have put two default gateway with different distance for
both subnet. For avoiding the access to non default gateway VIP, I
have put different set of DNS for each VIP. This way can push the
clients to always use VIP A if the DNS come to VIP A (lower distance
subnet) by using VIP A IP for all DNS records bound with VIP A and
vice versa. If DNS query is coming to VIP B (higher distance subnet),
the reply will never acknowledged by the clients since it will be
routed through the VIP A. This way can make only VIP A will do the
good reply as long the ISP A link is good.
Best regards,
On Sun, Dec 13, 2009 at 8:18 AM, Jack Stewart <[email protected]>
wrote:
Hi Lazuardi,
You are correct in saying that you need VIP's on both subnets. Only a VIP
subnet A can use ISP A and only a VIP on Subnet B can use ISP B.
With ISP A as your default route, VIPs on Subnet A work just fine. The
load
balancer sees the destination address of the client and sends it to the
default route. But when the real server sends return traffic for a
traffic
that came in from VIP B, the routing table tells it to go to the gateway
for ISP A. Somewhere along the path the return traffic will get dropped
into
the bit bucket. A route-map can fix this.
A route-map with the "ip policy route-map" command applies a set of ACL's
and route related actions to traffic before the load balance looks at the
routing table. Route-maps have precedence. In an earlier example below, I
took advantage of using an extended ACL in order to force the return
traffic
for a VIP on subnet B to go out through the Gateway B. The load balancer
changes the source address of the real server to the VIP, my route map
saw
that the source address of the VIP was on subnet B, and so it set the
next
hop to Gateway B. The destination address and default gateway did not
matter
for this traffic.
This is a pretty terse description, let me know if it is along the right
track and we can get into details off line if needed.
The web is a can of worms. Is the content dynamic or static? Is the
traffic
encrypted or is it not encrypted? There are a lot of different tricks. I
would start with the basics - don't worry about it unless it is a problem
or
you know will be a problem. For the most part the client will cache the
destination IP for a period of time and most web content is static. Also,
are we talking encrypted or unencrypted traffic? If it is encrypted you
can
go off of the ssl session id. If this is really a problem, I'll put on my
sysadmin thinking cap but I'll need to know O/S, http. For example,
Apache
2.2 has a mod_header module so that you can set the particular server.
This
only applies to unencrypted traffic as the load balancer can't decrypt
it.
Lazuardi Nasution wrote:
Hi Jack,
I'm interested with route-maps. Can you tell me more about that ?
Currently I just using static route for two default gateway with
different distance.
My requirements is simple, just to do server load balancing no matter
where the traffic come in. Since I have two ISPs which give me two
subnets, so I think I must have two VIPs on different subnets too. I
need Cookie Switching since some of clients are behind the proxy or
NAT. I cannot use Source NAT since I have to preserve clients source
IP for logging purpose on the real server. Since I don't have to do
the link load balancing, I do static routing for default gateway, so
only single link will be bidirectional on the same time. But I don't
have any idea how to make those VIPs share the session table or at
least threat the request not as first request if there is related
cookie inside (ex. ServerID). If I can do that I'm sure I can switch
the request to the same real server no matter which VIP the request
come in.
Any idea ?
Best regards,
On Fri, Dec 11, 2009 at 4:19 AM, Jack Stewart <[email protected]>
wrote:
Hi Lazuardi,
I'm running the routing code. To the best of my knowledge, route-maps
apply
only to the routing code.
route-maps allow you to set the gateway by source address, destination
address, or port #. Static routes allow you to set gateway by
destination.
The VIP is the source address of the outgoing traffic. So in the
example
below, gateway for a VIP is based on its address (and not the
destination
client). This was a really hard concept for me to wrap my head around.
It isn't clear to me that your case is the same. My setup is very
atypical.
I had a lot of trouble debugging it (traffic would leave the client but
never come back to it). You might be able to get at it by looking at
the
interface traffic of the gateways.
Let me know the solution you come up with, I'm curious.
---Jack
Lazuardi Nasution wrote:
Hi Jack,
I think this solution is for Switch Code since with Router Code I can
have many Management IP even with different subnets. The default
gateway can be specified statically on the routing table or by using
routing protocol from the routers.
Best regards,
On Fri, Dec 11, 2009 at 2:25 AM, Jack Stewart <[email protected]>
wrote:
Hi Lazuardi,
I ran into similar issues - this is what ultimately work in my
environment.
It may not be the same but hopefully there are some takeaways. Please
let
me
know how much of this makes sense - that feedback will be helpful
with
my
documentation.
First, DNS is special and the following is generic.
You can only have one management IP and one default route. The
management
IP
should live on the same subnet that has the default route. The first
public
subnet with the management IP & default route (Pub_Subnet_1) was not
a
problem.
All of my real servers are on a different private subnet that the two
public
subnets and they all have the Load Balancer defined as their default
gateway.
To get subnet 2 (pub_subnet_2), I needed to define a router interface
for
that subnetwork (ve2) and policy routing/route-maps. The route-maps
are
for
making sure that the return traffic goes out via the same gateway
that
it
came in for non directly attached subnets. The way the mapping works
for
me
in the configuration is:
!
ip access-list extended match_pub_subnet_2
permit ip match_pubsub2/24 any
!
route-map more_default_routes permit 10
match ip address match_subnet_2
set ip next-hop subnet_2_gateway
!
ip policy prefer-direct-route
ip policy route-map more_default_routes
In route-maps, the 'permit #' is just the precedence order. You can
add
additional entries to a route-map. Route-maps are processed before
static
routes.
Lastly, I defined an outside NAT policy on Public_Subnet_1 for
traffic
originating private subnet traffic (i.e. directly attached servers).
I'm
not
100% sure this is a requirement but it helps with traceroute, etc.
In my case it was necessary to add VRRP but that is because I've more
than
one box and it isn't clear you need that.
Once this was done, everything worked nicely from outside to inside.
This is a global static approach. Most people seem to route-maps to
filter
routing protocols, but I'm not allowed to exchange LB routing
protocols
with
our routers by policy.
For VIPs and real servers on the same private subnet, I found that
either
DSR or source-nat with ACL's works well. If you are using DSR with
Linux
(it
seems to apply to other 2.6 kernels), you'll probably want to look at
the
brocade wiki).
With DNS, source-nat with ACL's is probably the simplest and easier
way
to
go.
---Jack
Lazuardi Nasution wrote:
Hi Jack,
Yes, there is different router per subnet and I have done the static
routing for that. VIP1 is in the same subnet with Management IP and
the Router1 is connected to eth1, so I just simply put Management IP
on eth1. Since Router2 is connected to eth2, should I do something
on
eth2, ex. put another management IP on the eth2 which is in the same
subnet with VIP2 ? The other ethernet ports are for Real Server so I
have give ve1 for those ports.
There is another weird problem. I have made DNS binding from VIP1
and
RE1 and I have put ve1 IP in the same subnet with RE1. RE1 default
gateway is ve1 IP. I can query the DNS through VIP1 but RE1 cannot
do
traceroute to the Internet, stuck on the ServerIron. What's happen
here ?
Best regards,
On Thu, Dec 10, 2009 at 3:41 AM, Jack Stewart <[email protected]>
wrote:
Hi Lazuardi,
Yeah! A question that might be up my alley. I've done this however
I
need
some more details.
Do these VIPs need different "static" default gateways on a per
subnet
basis? It's possible with the routing code and I can send out the
details
if
you are interested.
Otherwise the main trick with subnet A to subnet B traffic is to
make
sure
that the return traffic goes though the load balancer. The client &
server
need to see the Load Balancer as the gateway between subnet A &
subnet
B.
DSR and source NAT are also options.
So more details, please. Depending on what you need to do it might
help
knock out some of my documentation.
---Jack
Lazuardi Nasution wrote:
Hi,
Is it possible to have multiple VIPs with different Subnets on
ServerIron 4G or ServerIron ADX1000 ? How can I do that ? I'm
using
router code of firmware.
--
Jack Stewart
Academic Computing Services, IMSS,
California Institute of Technology
[email protected]
626-395-4690 office / 626-437-6035 cell
--
Jack Stewart
Academic Computing Services, IMSS,
California Institute of Technology
[email protected]
626-395-4690 office / 626-437-6035 cell
_______________________________________________
foundry-nsp mailing list
[email protected]
http://puck.nether.net/mailman/listinfo/foundry-nsp
