Here is one source of information on distributed data centers:
http://www.cisco.com/networkers/nw02/post/presentations/docs/NSC-284.pdf
-
You have a few options for doing this, but some is dependent on how your
network is setup:
- How do your data centers see each other (Layers 1/2/3)?
Are they IBGP full-mesh, or at least route-reflector/confederation?
How is OSPF connected (single area, multiple area, NSSA)?
What routes are where (static to BGP, static to OSPF, redistribution, etc)?
Are your data centers connected with a transparent LAN?
- How is your BGP setup (especially Inter-Domain or externally)?
Where does the IP space come from (traditional classful networks i.e.
0/1, 128/2, 192/3)?
How are your annoucements seen according to different sources (e.g.
Internet routing table, RIR's like ARIN, IRR's like RADB)?
Who do you multi-home and multi-link to and at which data center or both?
What is the use of LSR/MSR's (less and more specifics) and BGP attributes
(weight/local-pref, AS-path prepends/pollution, MED/metric, communities
especially upstream e.g. no-export, selective prepends, setting local-pref)?
-
Here are the options for distributed data centers or global load
balancing (GLB):
1) HTTP routing (typically depends on option 2 by way of DNS insertion)
Product choices: Cisco Content Router 4400-B, CSS WebNS proximity
Idea: use of Layer 4 information or HTTP/302 defined redirects
Broken: browser problems, DNS delegation problems (need for every
site), users can bypass redirection by going directly to previously
visited IP/DNS URL (e.g. using bookmarks), only works for HTTP
(not SMTP or other important applications that require load balancing)
2) DNS routing
Product choices: Cisco Distributed Director, CSS w/o proximity, Content
Router 4450, and DNS Round-Robin (multiple IN A resource records)
Idea: DNS round-robin but smarter, can be combined with ideas like
edge-intercept where you administratively control the end-user DNS.
Can also be combined with smart site-selection techniques like
using DD IGP/BGP/RTT, Boomerang's one-way race metric and RTT's,
and the CR4450 Scouting or E2SS feature/functionality (see the ICDN
and ECDN documentation or the CECOC course with CSS WebNS)
Broken: Proximity methods are better (end-user DNS != end-user), and
when DNS server caching cannot be controlled (e.g. the global Internet)
then DNS administrators can ignore low TTL's (and generally do)
3) Anycast
Products : Cisco RHI with Catatalyst 6k CSM or IOS SLB
Idea: announce a /32 host route and inject it into OSPF or BGP with a
metric. When service degrades or becomes unavailable, remove route.
Implement at least two servers at two different locations with same IP
Problems: System logs (like web logs) are split up across multiple places,
however this idea scales really well if you understand how it works
4) ARP/subnet take-over
Products : Cisco ONS 15454 TLS + CSS redundancy or CSM FT-VLAN
Idea: Fail over from one location to another by clearing the arp-cache at
the location where the data moves from and re-creating the subnet or
IP addresses at the new location where the data needs to be moved to
Problems: Works really well, but difficult to scale past two distributed sites.
This technique is typically only done with active/passive redundancy
-
If you think I missed something, please let me know.
Andre
