I previously described how to configure HA Proxy and ucarp to load balance traffic between servers and share common IP addresses in the event of failure, however this still leaves holes in the availability of the application. The scenario only accounts for availability within a single data center and does not address how traffic and application services are managed between two or more geographically distinct data centers.

I’ve attached a more simplified diagram of our current architecture to help highlight single points of failure (Bytemarks load balancing was simplified in the diagram).

It becomes much more complicated to handle fail-over between multiple data centers. As an example if data center 1 fails entirely, we need to ensure that VIPs are routed to the correct data center OR DNS is changed. This becomes a two fold problem, by the time your DNS is propagated there is an unknown amount of time before becoming available again, in addition if you do not own the IP you cannot port to a different data-center.

There are a number of considerations that you can take into account, each will have varying implications on cost.

DNS

There are several things you can do at the DNS level to reduce the effect of any outage.

  • Have a very low time to live, or TTL in your DNS
  • If you have complete replica of your environment, split the traffic between the data centers using a round robin DNS (Will half the outage to those cached on the active data center)
    • Browsers will cache the DNS for 30 minutes
    • ISPs and Corporates will cache your DNS, regardless of TTL
  • Maintain a minimal setup (1 or 2 nodes) to balance traffic between the data-centers (using round robin as described above) and utilize HA Proxy weighted to a data center. In the event of failure more nodes can be setup automatically to recover.

Multihoming via routing

All of the above still leaves margin for outages and DNS can not be used to ensure high availability alone, despite helping to some degree. As mentioned above it is possible to move the network addresses from one data center to another data center. However re-routing IP addresses becomes fairly tricky if you are working with small CIDR blocks and are specific to the IP Network Allocation used. In addition it involves coordination with the hosting provider and upstream transit providers.

There are two types of IP Network Allocation.

  • PA – Provider Agregable
  • PI – Provider Independent

PA – Provider Agregable

  • Ripe assigns a large block of networks to a LIR (Local Internet Registry)
  • LIR assigns smaller networks to customers from the larger block
  • PA addresses cannot be transferred between providers
  • Request form to be filled by each end-customer justifying quantity

PI – Provider Independent

  • Not associated with any provider –cannot be agregated
  • Used by some dual-homed networks
  • RIPE performs much stricter checking of application than for PA
  • Applicant must justify “why not PA?”
  • Smallest allocation /24 (255 IP addrs)
  • LIR (Local Internet Registry) submits form on customer’s behalf

IP Addressing and Registrars Explained

In order to have high availability and re-route traffic you will need the following.

  • Your own address block (or a block of ‘provider independent’ space)
  • Your own Autonomous System number, and
  • Multiple upstream network connections.

Whilst it is possible to use Provider Agregable addresses and advertise the fragment to the other providers. “Wherever the fragment is propagated in the Internet the incoming traffic will follow the path of the more specific fragment, rather than the path defined by the aggregate announcement” and therefore will require Provider Independent addresses.

In order to acquire PI addresses you must register them through either an LIR (Local Internet Registry) or by becoming an LIR through a Regional Internet Registry (RIR).

Registry Area Covered
AfriNIC Africa Region
APNIC Asia/Pacific Region
ARIN North America Region
LACNIC Latin America and some Caribbean Islands
RIPE NCC Europe, the Middle East, and Central Asia

It is relatively cheap to acquire AS numbers and PI addresses, which can be acquired through Secura Hosting, which is an LIR.

Once you have your address block, AS number and multiple upstream connections, you announce your address block to each provider, and receive their routing table via an eBGP session. You can then configure quagga a software routing suite to do this.

For more information on BGP see:
BGP Blackmagic: Load Balancing in “The Cloud”

Setting up Quagga

su
apt-get install quagga quagga-doc
touch /etc/quagga/{bgpd.conf,ospfd.conf,zebra.conf}
sed -i 's/bgpd=no/bgpd=yes/' /etc/quagga/daemons
chown quagga.quaggavty /etc/quagga/*.conf 

echo 'password YourPassHere' > /etc/quagga/bgpd.conf
echo 'password YourPassHere' > /etc/quagga/ospfd.conf
echo 'password YourPassHere' > /etc/quagga/zebra.conf

sed -i 's/bgpd=no/bgpd=yes/' /etc/quagga/daemons

# By default Linux block traffic that go out from an interface and come back from another.
sed -i 's/.net.ipv4.conf.all.rp_filter=[0,1]/net.ipv4.conf.all.rp_filter=0/;s/.net.ipv4.conf.lo.rp_filter=[0,1]/net.ipv4.conf.lo.rp_filter=0/;s/.net.ipv4.conf.default.rp_filter=[0,1]/net.ipv4.conf.default.rp_filter=0/' /etc/sysctl.conf
sysctl -p

/etc/init.d/quagga start

Configuring Quagga for BGP

andrew-home:~# telnet localhost bgpd
Trying ::1...
Trying 127.0.0.1...
Connected to localhost.
Escape character is '^]'.

Hello, this is Quagga (version 0.99.15).
Copyright 1996-2005 Kunihiro Ishiguro, et al.

User Access Verification

Password:
andrew-home>
andrew-home> enable
andrew-home# conf t
andrew-home(config)# hostname R1
R1(config)# router bgp 65000
R1(config-router)# network 10.1.2.0/24
R1(config-router)# ^Z
R1# wr
Configuration saved to /usr/local/quagga/bgpd.conf
R1# show ip bgp
BGP table version is 0, local router ID is 0.0.0.0
Status codes: s suppressed, d damped, h history, * valid, > best, i - internal,
              r RIB-failure, S Stale, R Removed
Origin codes: i - IGP, e - EGP, ? - incomplete

   Network          Next Hop            Metric LocPrf Weight Path
*> 10.1.2.0/24      0.0.0.0                  0         32768 i

Total number of prefixes 1
R1# exit
Connection closed by foreign host.

# A few addition commands
show ip bgp summary

I wont go too much into configuring this, however there there a few additional resources that can help.

All transit providers for the site accept a prefix advertisement from the multi-homed site, and advertise this prefix globally in the inter-domain routing table. When connectivity between the local site and an individual transit provider is lost, normal operation of the routing protocol will ensure that the routing advertisement corresponding to this particular path will be withdrawn from the routing system, and those remote domain domains who had selected this path as the best available will select another candidate path as the best path. Upon restoration of the path, the path is re-advertised in the inter-domain routing system. Remote domains will undertake a further selection of the best path based on this re-advertised reachability information. Neither the local or the remote host need to have multiple addresses, nor undertake any form of address selection.
Multi-Homing

Problems

This also does have its problems for Provider Independent addresses, which causes concern for this approach to multi-homing via Provider Independent addresses. However Provider Agregable addresses do not typically suffer from the first two points below.

  • RIPE warns: “assignment of address space does NOT imply that this address space will be ROUTABLE ON ANY PART OF THE INTERNET”, see PA vs. PI Address Space
    • Possible auto summarization on global routing tables. Routing protocols summarize multiple routes into single one to cut down size of the routing tables… So assuming your class is A.B.C.D/26, you may have one or more ISPs having summary route to A.B.0.0/16 pointing to completely different network then your IP is physically in. Understanding Route Aggregation in BGP
  • Small subnets such as /24 are heavily penalized by BGP dampening after few flaps.
    • “When starting out with BGP, opening sessions with other providers, if you have got tweaking to do and it does start flapping, some providers will simply disconnect you, whilst others will dampen for days. Basically closing any throughput, watching and waiting to see if you stabilize. Others will require require contracts to be drawn up to prove you have competency and proven stability.”, Andrew Gearing
  • Many ISPs filter the IP address routing announcements made by their customers to prevent malicious activities such as prefix hijacking. As a result, it can take time for ISPs to reconfigure these filters which are often manually maintained. This can slow down the process of moving IP addresses from the primary datacenter to the backup datacenter.
  • Smaller routes won’t be allowed on the global backbone routing BGP tables, because they don’t want a huge table of small allocations – for performance and stability the backbone prefers a small number of relatively large address space chunks.

The above points do not really make this approach feasible for small CIDR blocks, as it will be difficult to justify a large enough range of routes as well as ensuring infrastructure for BGP.

High Availability through EC2

We originally hosted a large number of servers with EC2, however have moved most of our services away from EC2 due to excessive costs with large and extra large instances. At present we have moved a large number to bytemark. The following does put a dependency on a single provider…

Implementation

  • EC2 Elastic Load balancing
  • Two small EC2 compute nodes in different data centers/region soley for proxying content
  • HA Proxy

Benefits:

  • Using Elastic Load Balancing, you can distribute incoming traffic across your Amazon EC2 instances in a single Availability Zone or multiple Availability Zones.
  • Elastic Load Balancing can detect the health of Amazon EC2 instances. When it detects unhealthy load-balanced Amazon EC2 instances, it no longer routes traffic to those Amazon EC2 instances instead spreading the load across the remaining healthy Amazon EC2 instances.
    • Trivial to monitor and spawn instances on failure and re-assign IP addresses
  • No need to configure BGP and acquire Provider Independent addresses
  • Cost, bandwidth and the elastic load balancer are relatively cheap. ” transferring 100 GB of data over a 30 day period, the monthly charge would amount to $18 (or $0.025 per hour x 24 hours per day x 30 days x 1 Elastic Load Balancer) for the Elastic Load Balancer hours and $0.80 (or $0.008 per GB x 100 GB) for the data transferred through the Elastic Load Balancer, for a total monthly charge of $18.80.”. Small EC2 instances are fairly cheap, and solely forwarding requests.
  • EC2 Auto Scaling can be used to spawn more nodes and configure automatically.

Disadvantages:

  • “The ELB system strives to scale ahead of demand but wild spikes in traffic demand (such as 0 load to full load swings during initial load testing) can temporarily run past it’s provisioned capacity which will result in 503 Service Unavailable until the system can get ahead of demand. Under extreme overload, timeouts can occur. Customer traffic flows are typically gradual swings that occur at rates measured in minutes or hours, not fractions of a second.”
    • This is a problem for load testing, as the ELB will drop connections, however ELB as described will increase capacity gradually, to keep up with demand. Will need to run some load tests to test this, and check whether the service can adequately handle our traffic needs.
  • Elastic Load Balancer: An Elasticity Gotcha (Caching DNS and ELB releasing IPs) and Elastic Load Balancer is elastic – how the elasticity is achieved
    • ELB does not immediately release any IP address when traffic decreases. Instead, it removes the IP address from the associated DNS name. However, the IP address continues to be associated with the load balancer for some period of time. During this time, ELB monitors various metrics to check if this IP address is still receiving traffic, and release it when it is appropriate to do so.

      , Ameet@AWS

  • Cannot load balance via ELB between regions.
  • Health check via ELB minimum is 5 seconds, (Easy to monitor availability between nodes and remove from ELB ).
  • Additional latency, however initial tests seem neglible.
# WARNING: changing the keypair associated to an existing key will and running
# ./load.balancing.sh remove will delete all associated instances. Probably best to leave as is.
KEYPAIR=loadbalancing-test;
ELB_NAME='test';
NODES=2;
AMI=ami-b8446fcc;

function remove_elb
{
  #elb-delete-lb $ELB_NAME;
  ec2-describe-instances | grep $KEYPAIR | grep -v terminated | awk '{print $2}' | while read instanceId; do ec2-terminate-instances $instanceId; done;

  ec2-delete-keypair $KEYPAIR && rm ~/.ssh/$KEYPAIR;
}

function create_elb
{
  ec2-add-keypair $KEYPAIR > ~/.ssh/$KEYPAIR;
  chown 0600 ~/.ssh/$KEYPAIR;
  ssh-keygen -p -f ~/.ssh/$KEYPAIR;

  elb-create-lb $ELB_NAME --availability-zones $EC2_AVAILABILITY_ZONE --listener "protocol=http,lb-port=80,instance-port=80"

  # Create instances to attach to ELB
  for i in `seq 1 $NODES`; do
    ec2-run-instances $AMI --user-data-file install-lamp -k $KEYPAIR
  done;

  # Authorize CIDR Block 0.0.0.0/0
  ec2-authorize default -p 22
  ec2-authorize default -p 80

  addresses_allocated=0;
  ec2-describe-instances | grep $KEYPAIR | grep -v terminated | awk '{print $2}' | while read instanceId; do
    echo elb-register-instances-with-lb $ELB_NAME --instances $instanceId;
    elb-register-instances-with-lb $ELB_NAME --instances $instanceId;

#    # Allocate addresses for each node
#    # You may need to contact EC2 to increase the amount of IP addresses you can allocate, however these will be hidden, so not important
#    while [ "`ec2-describe-addresses | grep -v 'i-' | awk '{print $2}' | wc -l`" -lt "$NODES" ]; do
#      ec2-allocate-address;
#    fi;
#
#    # Allocate non associated addresses
#    ec2-describe-addresses | grep -v 'i-' | awk '{print $2}' | while read ip_address; do
#      echo $ip_address;
#      ec2-associate-address $ip_address -i $instanceId
#      addresses_allocated=$(($addresses_allocated+1));
#    done;

  done;

  elb-configure-healthcheck $ELB_NAME  --headers --target "TCP:80" --interval 5 --timeout 2 --unhealthy-threshold 2 --healthy-threshold 2
  elb-enable-zones-for-lb $ELB_NAME --availability-zones $EC2_AVAILABILITY_ZONE

}

function test_elb
{
  SERVER="`ec2-describe-instances | grep $KEYPAIR | grep -v terminated | head -n1 | awk '{print $4}'`"
  instanceId="`ec2-describe-instances | grep $KEYPAIR | grep -v terminated | head -n1 | awk '{print $2}'`"
  echo "Shutting down $SERVER";
  ssh -i ~/.ssh/$KEYPAIR root@$SERVER "/etc/init.d/apache2 stop"

  sleep 11;
  elb-describe-instance-health $ELB_NAME

  ssh -i ~/.ssh/$KEYPAIR root@$SERVER '/etc/init.d/apache2 start'
  sleep 2;
  elb-describe-instance-health $ELB_NAME
}

if ( [ $# -gt 0 ] && ( [ $1 = "create" ] || [ $1 = "remove" ] || [ $1 = "test" ] || [ $1 = "health" ] ) ); then
  command="$1;"
else
  echo "Usage `basename $0` create|remove";
  echo "Usage `basename $0` test";
  exit;
fi;

case $1 in
  "remove")
    remove_elb;
  ;;
  "create")
      create_elb;
  ;;
  "test")
      test_elb;
  ;;
  "health")
        elb-describe-instance-health $ELB_NAME;
  ;;
esac;

Node configuration:

Change the ip addresses 192.0.32.10:80 to the relevant servers, presently this points to example.com.

install-lamp

#!/bin/bash

export DEBIAN_FRONTEND=noninteractive;
apt-get update && apt-get upgrade -y && apt-get install -y apache2 haproxy subversion;
tasksel install lamp-server;
echo "Please remember to set the MySQL root password!";
IP_ADDRESS="`ifconfig  | grep 'inet addr:'| grep -v '127.0.0.1' | cut -d: -f2 | awk '{ print $1}'`";
echo $IP_ADDRESS >/var/www/index.html;

cat > /etc/haproxy/haproxy.cfg < < EOF
global
	log /dev/log    local1
	maxconn 4096
	#chroot /usr/share/haproxy
	user haproxy
	group haproxy
	daemon
	#debug
	#quiet
defaults
	log     global
	mode    http
	option  httplog
	option  dontlognull
	retries 3
	option redispatch
	maxconn 2000
	contimeout      5000
	clitimeout      50000
	srvtimeout      50000
	option      httpclose

listen gkweb $IP_ADDRESS:80
	mode http
	balance leastconn
	stats enable
	stats realm LB Statistics
	stats auth stats:password
	stats scope .
	stats uri /stats?stats
	#persist
	server web1 192.0.32.10:80 check inter 2000 fall 3
	server web2 192.0.32.10:80 check inter 2000 fall 3
	server web3 192.0.32.10:80 check inter 2000 fall 3
	server web4 192.0.32.10:80 check inter 2000 fall 3
EOF

sed -i -e 's/^ENABLED.*$/ENABLED=1/' /etc/default/haproxy;
/etc/init.d/apache2 stop;
/etc/init.d/haproxy restart;

Commercial:

References:
Multihoming to one ISP
Multipath Load Sharing with two ISPs
Small site multihoming
EBGP load balancing with EBGP session between loopback interfaces
Failover and Global Server Load Balancing for Better Network Availability

The “Elastic” in “Elastic Load Balancing”: ELB Elasticity and How to Test it
EC2 Instance Belonging to Multiple ELBs, (Creating different tiers of service, Splitting traffic across multiple destination ports)