Linux Virtual Server Tutorial

Horms (Simon Horman) - horms@valinux.co.jp

VA Linux Systems Japan, K.K. - www.valinux.co.jp

July 2003. Revised March 2004



http://www.ultramonkey.org/

with assistance from

Abstract:

The Linux Virtual Server Project (LVS) allows load balancing of networked services such as web and mail servers using Layer 4 Switching. It is extremely fast and allows such services to be scaled to service 10s or 100s of thousands of simultaneous connections. The purpose of this tutorial is to demonstrate how to use various features of LVS to load balance Internet services, and how this can be made highly available using tools such as such as heartbeat and keepalived. It will also cover more advanced topics which have been the subject of recent development including maintaining active connections in a highly available environment and using active feedback to better distribute load.

Introduction

The Linux Virtual Server Project (LVS) implements layer 4 switching in the Linux Kernel. This allows TCP and UDP sessions to to be load balanced between multiple real servers. Thus it provides a way to scale Internet services beyond a single host. HTTP and HTTPS traffic for the World Wide Web is probably the most common use. Though it can also be used for more or less any service, from email to the X Windows System.

LVS itself runs on Linux, however it is able to load balance connections from end users running any operating system to real servers running any operating system. As long as the connections use TCP or UDP, LVS can be used.

LVS is very high performance. It is able to handle upwards of 100,000 simultaneous connections. It is easily able to load balance a saturated 100Mbit ethernet link using inexpensive commodity hardware. It is also able to load balance saturated 1Gbit link and beyond using higher-end commodity hardware.

LVS Basics

This section will cover the basics of how LVS works. How to obtain and install LVS, and how to configure for its main modes of operation. In short it will cover how to set up LVS to load balance TCP and UDP services.

Terminology

Linux Director: Host with Linux and LVS installed which receives packets from end users and forwards them to real servers.

End User: Host that originates a connection.

Real Server: Host that terminates a connection. This will be running some sort of daemon such as Apache.

A single host may be act in more than one of the above roles at the same time.

Virtual IP Address (VIP): The IP address assigned to a service that a Linux Director will handle.

Real IP Address (RIP): The IP address of a Real Server.

Layer 4 Switching

Figure 1: LVS NAT
\includegraphics[width=15cm]{lvs-nat}

Layer 4 Switching works by multiplexing incoming TCP/IP connections and UDP/IP datagrams to real servers. Packets are received by a Linux Director and a decision is made as to which real server to foward the packet to. Once this decision is made subsequent packets to for the same connection will be sent to the same real server. Thus, the integrity of the connection is maintained.

Forwarding Packets

The Linux Virtual Server has three different ways of forwarding packets; network address translation (NAT), IP-IP encapsulation (tunnelling) and direct routing.

Figure 2: LVS Direct Routing
\includegraphics[width=15cm]{lvs-dr}

Virtual Services

On the Linux Director a virtual service is defined by either an IP address, port and protocol, or a firewall-mark. A virtual service may optionally have a persistance timeout associated with it. If this is set and a connection is received from the same IP address before the timeout has expired, then the connection will be forwarded to the same real server as the original connection.

Scheduling

The virtual service is assigned a scheduling algorithm that is used to allocate incoming connections to the real servers. In LVS the schedulers are implemented as separate kernel modules. Thus new schedulers can be implemented without modifying the core LVS code.

There are many different scheduling algorithms available to suit a variety of needs. The simplest are round robin and least connected. These work using a simple strategy of allocating connections to each real server in turn and allocating connections to the real server with the least number of connections respectively. Weighted variants of these schedulers allow connections to be allocated proportional to the weighting of the real server, more powerful real servers can be set with a higher weight and thus, will be allocated more connections.

More complex scheduling algorithms have been designed for specialised purposes. For instance to ensure that requests for the same IP address are sent to the same real server. This is useful when using LVS to load balance transparent proxies.

Installing LVS

Some distributions, such as SuSE ship with kernels that have LVS compiled in. In these cases installation should be as easy as installing the supplied ipvsadm package. At the time of writing Ultra Monkey provides packages built against Debian Sid (Unstable) and Woody (Stable/3.0) and Red Hat 7.3 and 8.0. Detailed information on how to obtain and install these packages can be found on www.ultramonkey.org. The rest of this section will discuss how to install LVS from source as it is useful to understand how this process works.

Early versions of LVS worked with Linux 2.2 series kernels. This implementation involved extensive patching of the Kernel sources. Thus, each version of LVS was closely tied to a version of the Kernel. The netfilter packet filtering architecture[4] which is part of the 2.4 kernels has allowed LVS to be implemented almost exclusively as a set of kernel modules. The result is that LVS is no longer tied closely to an individual kernel release. LVS may also be compiled directly into the kernel. However, this discussion will focus on using LVS as a module as this approach is easier and more flexible.

  1. Obtain and Unpack Kernel

    It is always easiest to start with a fresh kernel. You can obtain this from www.kernel.org. This example will use the 2.4.20 kernel. It can be unpacked using the following command which should unpack the kernel into the linux-2.4.20 directory.

    tar -jxvf linux-2.4.20.tar.bz2
    

  2. Obtain and Unpack LVS

    LVS can be obtained from www.linuxvirtualserver.org. This example will use 1.0.9. It can be unpacked using the following command which should pack the kernel into the ipvs-1.0.9 directory.

    tar -zxvf ipvs-1.0.9.tar.gz
    

  3. Apply LVS Patches to Kernel

    Two minor kernel patches are required in order for the LVS modules to compile. To apply these patches use the following:

    cd linux-2.4.20/
    patch -pq < ../ipvs-1.0.9/linuxkernel_ksyms_c.diff
    patch -pq < ../ipvs-1.0.9/linuxnet_netsyms_c.diff
    

    A third patch is applied to allow interfaces to be hidden. Hidden interfaces do not respond to ARP requests and are used on real servers with LVS direct routing.

    patch -pq < ../ipvs-1.0.9/contrib/patches/hidden-2.4.20pre10-1.diff
    

  4. Configure the kernel

    First ensure that the tree is clean:

    make mrproper
    

    Now configure the kernel. There are a variety of ways of doing this including make menuconfig, make xconfig and make config. Regardless of the method that you use, be sure to compile in netfilter support, with at least the following options. It is suggested that where possible these options are built as modules.

    Networking options  --->
      Network packet filtering (replaces ipchains)
      <m> IP: tunnelling
      IP: Netfilter Configuration --->
         <m> Connection tracking (required for masq/NAT)
           <m> FTP protocol support
         <m> IP tables support (required for filtering/masq/NAT)
           <m> Packet filtering
             <m> REJECT target support
           <m> Full NAT
             <m> MASQUERADE target support
             <m> REDIRECT target support
             <m> NAT of local connections (READ HELP) (NEW)
           <m> Packet mangling
             <m> MARK target support
           <m> LOG target support
    

  5. Build and Install the Kernel

    As the kernel has been reconfigured the build dependencies need to be reconstructed.

    make dep
    

    The kernel and modules may now be build using:

    make bzImage modules
    

    To install the newly built modules and kernel run the following command. This should install the modules under /lib/modules/2.4.20/and the kernel in /boot/vmlinuz-2.4.20

    make install modules_install
    

  6. Update boot loader

    In the case of grub is used as the boot loader then a new entry should be added to /etc/grub.conf. This example assumes that the /boot partition is /dev/hda3. Existing entries in /etc/grub.conf should be used as a guide.

    title 2.4.20 LVS
            root (hd0,0)
            kernel /vmlinuz-2.4.20 ro root=/dev/hda3
    

    If the boot loader is lilo then a new entry should be added to /etc/lilo.conf. This example assumes that the / partition is /dev/hda2. Existing entries in /etc/lilo.conf should be used as a guide.

    image=/boot/vmlinuz-2.4.20
            label=2.4.20-lvs
            read-only
            root=/dev/hda2
    

    Once /etc/lilo.conf has been updated run lilo.

    lilo
    Added Linux-LVS *
    Added Linux
    Added LinuxOLD
    

  7. Reboot the system.

    At your boot loader's prompt be sure to boot the newly created kernel.

  8. Build and Install LVS

    The commands to build LVS should be run from the ipvs-1.0.9/ipvs/ directory. To build and install use the following commands. /kernel/source/linux-2.4.20 should be the root directory that the kernel was just built in.

    make KERNELSOURCE=/kernel/source/linux-2.4.20 all
    make KERNELSOURCE=/kernel/source/linux-2.4.20 modules_install
    

  9. Build and Install Ipvsadm

    Ipvsadm is the user-space tool that is used to configure LVS. The source can be found in the ipvs-1.0.9/ipvs/ipvsadm/ directory. To build and install use the following commands.

    make all
    make install
    

LVS NAT

LVS NAT is arguably the simplest way to configure LVS. Packets from real servers are received by the linux director and the destination IP address is rewritten to be one of the real servers. The return packets from the real server have their source IP address changed from that of the real server to the VIP.

Figure 3: LVS NAT Example
\includegraphics[width=8cm]{lvs-nat-ex}

Linux Director

Real Servers

Testing and Debugging

Testing can be done by connecting to 172.17.60.201:80 from outside the server network.

Running a packet tracing tool on the linux directors and real servers is very useful for debugging purposes. Many setup problems can be resolved by tracing the path of a connection and observing at which step packets fail to appear. Using Tcpdump will be discussed here as an example, there are variety of tools available for various operating systems.

The following trace shows a connection being opened by an end user 10.2.3.4 to the VIP 172.17.60.201 which is forwarded to the real server 192.168.6.5. It shows packets being received by the linux director and then forwarded to the real server and vice versa. Note that the packets forwarded to the real server still have the end user's ip address as the source address. The linux director only changes the destination IP address of the packet. Similarly replies from the real servers have the destination address set to that of the end user. The linux director only rewrites the source IP address of reply packets so that it is the VIP.

tcpdump -n -i any port 80
12:40:40.965499 10.2.3.4.34802 > 172.17.60.201.80: 
        S 2555236140:2555236140(0) win 5840 
        <mss 1460,sackOK,timestamp 16690997 0,nop,wscale 0>
12:40:40.967645 10.2.3.4.34802 > 192.168.6.5.80: 
        S 2555236140:2555236140(0) win 5840 
        <mss 1460,sackOK,timestamp 16690997 0,nop,wscale 0>
12:40:40.966976 192.168.6.5.80 > 10.2.3.4.34802: 
        S 2733565972:2733565972(0) ack 2555236141 win 5792 
        <mss 1460,sackOK,timestamp 128711091 16690997,nop,wscale 0> (DF)
12:40:40.968653 172.17.60.201.80 > 10.2.3.4.34802: 
        S 2733565972:2733565972(0) ack 2555236141 win 5792 
        <mss 1460,sackOK,timestamp 128711091 16690997,nop,wscale 0> (DF)
12:40:40.971241 10.2.3.4.34802 > 172.17.60.201.80: 
        . ack 1 win 5840 <nop,nop,timestamp 16690998 128711091>
12:40:40.971387 10.2.3.4.34802 > 192.168.6.5.80: 
        . ack 1 win 5840 <nop,nop,timestamp 16690998 128711091>
ctrl-c

ipvsadm -L -n can be used to show the number of active connections.

ipvsadm -L -n
IP Virtual Server version 1.0.9 (size=4096)
Prot LocalAddress:Port Scheduler Flags
  -> RemoteAddress:Port           Forward Weight ActiveConn InActConn
TCP  172.17.60.201:80 rr
  -> 192.168.6.5:80               Masq    1      7          3         
  -> 192.168.6.4:80               Masq    1      8          4

ipvsadm -L -stats will show the number of packets and bytes sent and received per second.

ipvsadm -L -n --stats
IP Virtual Server version 1.0.9 (size=4096)
Prot LocalAddress:Port               Conns   InPkts  OutPkts  InBytes OutBytes
  -> RemoteAddress:Port
TCP  172.17.60.201:80                  114     1716     1153   193740 112940
  -> 192.168.6.5:80                     57      821      567    94642  55842
  -> 192.168.6.4:80                     57      895      586    99098  57098

ipvsadm -L -rate will show the total number of packets and bytes sent and received.

ipvsadm -L -n --rate
IP Virtual Server version 1.0.9 (size=4096)
Prot LocalAddress:Port                 CPS    InPPS   OutPPS    InBPS OutBPS
  -> RemoteAddress:Port
TCP  172.17.60.201:80                   56      275      275    18739  41283
  -> 192.168.6.5:80                     28      137      137     9344  20634
  -> 192.168.6.4:80                     28      138      137     9395  20649

ipvsadm -L -zero will zero all the statistics counters.

LVS Direct Routing

Figure 4: LVS Direct Routing Example
\includegraphics[width=15cm]{lvs-dr-ex}

LVS Direct Routing works by forwarding packets, unchanged, to the MAC addresses of real servers. As the packet is unmodified the real servers need to be configured to accept traffic addressed to the VIP. This is most commonly done by using a hidden interface.

As the incoming packets are not modified by the linux director the return packets do not need to pass through the linux director. Thus, higher throughput can be obtained. It is also easier to load balance services for end users on the same local network as the return packets can be sent directly to the end user rather than forcing them to go through the linux director.

Linux Director

Real Servers

Testing and Debugging

Testing can be done by connecting to 172.17.60.201:80 from any network.

Debugging can be done using ipvadm and packet tracing as per LVS NAT. However, note that when the packets are forwarded no address translation takes place. Also note that as the return packets are not handled by LVS - they are sent directly to the end user by the real server - the outgoing packet and byte statistics will be zero.

LVS Tunnel

Figure 5: LVS Tunnel Example - Same Topology as the LVS Direct Routing Example
\includegraphics[width=15cm]{lvs-dr-ex}

LVS tunnelling works in a very similar manner to direct routing. The main difference is that packets are forwarded to the real servers using IP encapsulated in IP, rather than just sending a new ethernet frame. The main advantage of this is that real servers may be on a different network to the linux director.

Linux Director

Real Servers

Testing and Debugging

Testing can be done by connecting to 172.17.60.201:80 from any network. Debugging is as per LVS direct routing.

High Availability

LVS is an effective way to load balance networked services. Typically this means that several servers will act, as far as end-users are concerned, as if they were a single server. Unfortunately, the more servers that are in the system, the greater the chance that a single server will fail. Thus, it is important to make use of high availability techniques to ensure that the virtual service is maintained even if individual servers fail.

Heartbeat

Heartbeat be used to monitor a pair of linux directors and ensure that one of them owns the VIP at any given time. It works by each host periodically sending a heartbeat message. If no heartbeat message is received for a predetermined period of time then the host is considered to have failed. When this occurs resources can be taken over. Heartbeat has a modular design that allows arbitrary resources to be defined.

For the sake of this discussion we will be using an IP address as a resource. When fail-over occurs the IP address is obtained using a method known as IP address take-over. This works by the newly activated linux director sending gratuitous ARP packets for the VIP. All hosts on the network should receive these ARP packets and thus send subsequent packets for the VIP to the new linux director.

Heartbeat can be obtained from www.linux-ha.org. It can also be installed by using the packages provided or built from source using the following commands.

./ConfigureMe build
make
make install

Sample Configuration

Figure 6: Heartbeat Example
\includegraphics[width=16cm]{heartbeat}

Configuration is done using three files that can be found in /etc/ha.d.

LVS should be configured the same way on both linux directors. For this example the LVS tunnel configuration discussed earlier will be used. Direct Routing and NAT may also be used.

As the VIP, 172.17.60.201 is managed by heartbeat it should not be brought up on the linux directors by other means.

Heartbeat should be started on both linux directors. After a few moments the VIP should be brought up on whichever linux director is the master.

Testing and Debugging

Reboot the active linux director and observe that the other linux director takes over. You can examine the progress of the take-over by examining the logs sent to syslog, typically found in /var/log/messages. As nice_failback is on, the currently active linux director will now act as the master and when the failed linux director comes back online it will act as a standby.

Ipfail

The design of heartbeat is such that if any communication channel is available to a host, then it will be considered to be available. This is not always the desired behaviour. For example if a pair of hosts have links on the internal and external network, it may be desirable for fail-over to occur if either link fails on one host. After all it can no longer communicate route traffic between end users and the real servers.

Figure 7: Heartbeat without IPfail
\includegraphics[width=16cm]{ipfail}

The ipfail plug-in for heartbeat makes this possible by monitoring one or more external hosts known as a ping node. Typically this would be a router or the switch itself. The ping node is treated as a quorum device. That is, if a host cannot access a ping node, it is not eligible to hold any resources. Thus, if an interface fails on the active linux director, then one of the ping nodes should become unavailable and fail-over will occur.

Figure 8: Heartbeat with IPfail
\includegraphics[width=16cm]{ipfail2}

The ipfail module is shipped as part of heartbeat. Additional information is available from http://pheared.net/devel/c/ipfail/. In the long term this will be integrated into the heartbeat documentation.

Sample Configuration

To use ipfail with the heartbeat setup discussed previously, the following should be added to heartbeat's ha.cf file.

ping 172.17.0.254
respawn hacluster /usr/lib/heartbeat/ipfail

A ping directive should be added for each ping node. I have only defined one for the external network, as there are no suitable quorum devices on the internal network in this demonstration.

The respawn directive tells heartbeat to run /usr/lib/heartbeat/ipfail as user hacluster. To rerun it if it exits with a status other than 100, and to kill it when heartbeat exits.

After adding these options heartbeat needs to be restarted.

/etc/init.d/heartbeat restart

Testing and Debugging

Testing and debugging can be done as per Heartbeat itself.

Ldirectord

Heartbeat is used to monitor the health of linux directors. Ldirectord can be used to monitor the health of real servers and manipulates the LVS kernel table accordingly. Ldirectord and heartbeat are often used in tandem to create a high availability LVS cluster.

Ldirectord checks services on the real servers by connecting to them, making a known request and checking the result for a known string. Checks are provided for HTTP, HTTPS, FTP, IMAP, POP, SMTP, LDAP and NNTP. Additional checks can be added by modifying the code, which is usually quite straight forward. In fact many of the checks incorporated by ldirectord have been supplied as patches by users.

The check semantics above are known as a negotiate check. Another type of check, the connect check, simply checks to make sure a connection can be opened to the service on the real server. This is useful if there is not a check for the protocol supplied by ldirectord.

Sample Configuration

Ldirectord is configured using the ldirectord.cf file. It has global directives which either set global options, such as where to log errors to, or defaults for the virtual services. The virtual services encapsulate a virtual service provided by LVS. The virtual services contain the real servers which are checked.

# Global Directives
checktimeout=10
checkinterval=2
autoreload=no
logfile="local0"
quiescent=yes

# Virtual Server for HTTP
virtual=172.17.60.201:80
        fallback=127.0.0.1:80
        real=192.168.6.4:80 masq
        real=192.168.6.5:80 masq
        service=http
        request="index.html"
        receive="Test Page"
        scheduler=rr
        protocol=tcp
        checktype=negotiate

Ldirectord may be started by running the ldirectord command, the ldirectord init script or by adding it as a resource to heartbeat. There is no particular advantage to the latter as ldirectord can happily run on the master and stand-by linux directors at the same time.

Once ldirectord has started the LVS kernel table will be populated.

ipvsadm -L -n
IP Virtual Server version 1.0.7 (size=4096)
Prot LocalAddress:Port Scheduler Flags
  -> RemoteAddress:Port           Forward Weight ActiveConn InActConn
TCP  172.17.60.201:80 rr
  -> 192.168.6.5:80               Masq    1      0          0
  -> 192.168.6.4:80               Masq    1      0          0
  -> 127.0.0.1:80                 Local   0      0          0

By default ldirectord uses the quiescent feature of LVS to add and remove real servers. That is, when a real server is to be removed its weight is set to zero and it remains part of the virtual service. This has the effect that exiting connections to the real server may continue, but no new connections will be allocated. This is particularly useful for gracefully taking real servers offline. This behaviour can be changed to remove the real server from the virtual service by setting the global configuration option quiescent=no.

Testing and Debugging

Testing can be done by bringing the real servers up and down. By changing the contents of the known URL that is being requested such that it does not contain the expected string. By killing the daemon that serves end-users' requests. Or by powering down the host all together.

In each case ldirectord should update the LVS kernel table accordingly which can be examined using ipvsadm -L -n. Ldirectord also logs its activities, the configuration above sets these logs to be written to syslog, typically they will show up in /var/log/syslog.

For extra debugging information ldirectord can be run in debugging mode, in which case it will log verbosely to the terminal and will not detach from the terminal. This is done by using the -d command line option. This example starts ldirectord in debugging mode with the configuration file ldirectord.cf, which should be in /etc/ha.d/. Debugging can be terminated using ctrl-c.

ldirectord -d ldirectord.cf start

Keepalived

Keepalived provides an implementation of the VRRPv2 protocol which is specified in RFC 2338[1]. It is an alternative method of managing a VIP on a network so that it is owned by only one host at any given time. This can be used to switch between active and stand-by linux directors.

VRRPv2 works on a simple state engine. Hosts advertise their availability. The highest priority host wins the resource and advertises this fact. All other nodes then go into the backup state.

There is another implementation of VRRPv2 for Linux from http://off.net/ jme/vrrpd/. However, at the time of writing the keepalived implementation appears to be much more complete.

Keepalived also features service level monitoring of real servers and manipulates the LVS kernel table accordingly. The service tests that are implemented are:

It also provides an API to implement new checks.

The VRRPD and LVS/Health-Check features can be used individually or in combination.

Keepalived is available from keepalived.sourceforge.net. It compilation is quite straight forward using ./configure and make. A .spec file for Red Hat is also provided. Packages for Debian are available in the main Debian tree.

To configure keepalived /etc/keepalived/keepalived.conf should be modified. This file is divided up into sections.

Note that the VRRP implementation works on a master/slave system. So each vrrp_instance should be marked as a "MASTER" on one node and a "SLAVE" on the other nodes. During testing, it did not appear possible to configure keepalived to have behaviour analogous to heartbeat's nice_failback. That is a node will hold a resource until it fails, in which case another node will take it over until it in turn fails. It was also found that the slave nodes should be given a lower priority than the master to avoid spurious fail-overs.

Sample Configuration

For the sake of brevity, the example configuration files are in Appendix A.

To create the checksums for the configuration file, the genhash programme can be used. Genhash will connect to the server and request the URL. It will then produce a lot of output, showing you how the data that is being used to construct the checksum. The final line is the checksum which should be included in keepalived.conf. For example, to generate the hash for the URL http://192.168.6.5:80/ the following command is used.

genhash -s 192.168.6.5 -p 80 -u /
[ lots of output omitted ]
90bfbce6bc089a41f1fddca9aeaba452

To start keepalived run the keepalived daemon or init script. Messages are logged to syslog and typically can be found in /var/log/message. After a few moments the LVS kernel table should be populated on both machines. This can be inspected using ipvsadm.

ipvsadm -L -n
IP Virtual Server version 1.0.7 (size=4096)
Prot LocalAddress:Port Scheduler Flags
  -> RemoteAddress:Port           Forward Weight ActiveConn InActConn
TCP  172.17.60.201:80 lc
  -> 192.168.6.5:80               Masq    1      0          0   
  -> 192.168.6.4:80               Masq    1      0          0

On the master machine the virtual ip addresses should have been added. This can be checked using the ip command.

ip addr sh
[ lo: omitted ]
2: eth0: <BROADCAST,MULTICAST,UP> mtu 1500 qdisc pfifo\_fast qlen 100
   link/ether 00:50:56:4f:30:19 brd ff:ff:ff:ff:ff:ff
   inet 172.17.60.207/16 brd 172.17.255.255 scope global eth0
   inet 172.17.60.201/32 scope global eth0
3: eth1: <BROADCAST,MULTICAST,UP> mtu 1500 qdisc pfifo\_fast qlen 100
   link/ether 00:50:56:4f:30:1a brd ff:ff:ff:ff:ff:ff
   inet 192.168.6.3/24 brd 192.168.6.255 scope
   global eth1 inet 192.168.6.1/32 scope global eth1

If a fail-over occurs the same addresses should appear on the slave, and then back on the master once it is restored.

New Developments

Active Feedback

Ldirectord, keepalived and other tools monitor the health of real servers. The weight parameter allows the relative capacity of real servers to be taken into account. However, these tools do not monitor the real-time serving capacity of the real servers and do not allocate connections proportional to this.

This can be particularly problematic in situations where some connections require significantly more resources on a real server than others. For instance, if some connections are a plain HTML file fetched from disk, or more likely memory. While other connections involve processing of information, such a scaling an image or retrieving part of the page from a database.

Feedbackd implements a framework that allows real-time information from from the real servers to determine how many connections they should be allocated relative to each other. As such, feedbackd implements an active feedback system. Feedbackd is available from http://www.redfishsoftware.com.au/projects/feedbackd/

Feedbackd has two key components, feedbackd-agent which runs on the real servers and monitors their serving capacity. The monitoring is modular so arbitrary checks can be defined. The default check supplied simply monitors CPU load using /proc/stat. The second component, feedbackd-master runs on the linux directors. It collates information from the feedbackd-agent's which connect and manipulates the weights of the real servers in the LVS kernel table accordingly.

It was found that a little bit of massaging was required to get it to compile. Also made minor enhancements were made to allow feedbackd-master to be restarted without giving "address in use" errors and to allow feedbackd-agent to timeout the master. The latter is a work around to allow feedbackd to work with Active/Stand-By Linux Directors. Both of these changes have been forwarded to the author and will hopefully show up in the next version.

The only configuration required for feedbackd-master is to establish the LVS virtual services that will be used. This is done using ipvsadm. There is no need to add the real servers as this will be done by feedbackd-master by matching the protocol and port information sent by the feedback-agents running on real servers. As such feedbackd can be used to add and remove real servers on the fly without any configuration of the linux director. For example:

ipvsadm -A -t 172.17.60.201:80

To start feedbackd-master simply run the daemon on the command line. No init script is supplied with the current distribution.

Feedbackd-Agent is configured by modifying /etc/feedbackd-agent.conf. In this file the Linux Director running feedbackd-master is specified as are the services that the real server should join.

director = 192.168.6.1
service = http
        protocol = TCP
        port = 80
        module = cpuload.so
        forwarding = NAT

Again, to run feedbackd-agent simply run the command on the command-line.

Testing

As a primitive test, one of the real servers can be loaded manually and the effects of this on the LVS table on the linux director can be observed using ipvsadm. An in-depth analysis of the effects of using feedbackd can be found in Jeremy Kerr's paper on the feedbackd[3].

Connection Synchronisation - Existing Solution

Configuring two linux directors in an active/stand-by configuration is a useful way to provide high availability. If the active linux director fails, the stand-by can automatically take over the IP address of the virtual services and the cluster can continue to function. However, when such a fail-over occurs connections that are currently in progress are terminated.

This is because the stand-by linux director does now know anything about these connections. By synchronising connection information between the active and stand-by linux directors this problem can be averted. Thus, when a stand-by linux director becomes the active linux director, it will have information about the currently active connections and will be able to continue to forward their packets. The critical piece of information required is which real server to forward packets for a given connection to. This information is quite small and thus can be synchronised with little overhead.

There is an implementation of connection synchronisation within the current LVS code. It works on a master/slave system whereby the linux director configured as the master sends synchronisation information for connections. The linux directors configured as slaves receive this information and update their LVS connection table accordingly.

A connection is synchronised once the number of packets passes a threshold (3) and then every frequency (50) packets. The synchronisation information for the connections are added to a queue and periodically flushed. The synchronisation information for up to 50 connections can be packed into a single packet. The packets are sent to the slaves using multicast.

Sending and receiving synchronisation information by the master and slaves respectively is done by a kernel thread. The kernel synchronisation thread is started on the master and slaves using the following commands.

ipvsadm --start-daemon master    # Run on the Master Linux Director
ipvsadm --start-daemon backup    # Run on the Slave Linux Director

Testing and Debugging

The synchronisation of connections can be monitored using ipvsadm -L -c -n, which lists LVS connection table. Connections should first appear on the master linux director. Then after a few moments, when synchronisation has occurs, they should also appear on the slaves.

ipvsadm -L -c -n # On the Master Linux Director
IPVS connection entries
pro expire state       source             virtual          destination
TCP 01:00  TIME_WAIT   172.16.4.222:34939 172.17.60.201:80 192.168.6.5:80
TCP 01:01  TIME_WAIT   172.16.4.222:34940 172.17.60.201:80 192.168.6.4:80
TCP 15:00  ESTABLISHED 172.16.4.222:34941 172.17.60.201:80 192.168.6.5:80

ipvsadm -L -c -n # On the Slave Linux Director
IPVS connection entries
pro expire state      source             virtual          destination
TCP 01.20 ESTABLISHED 172.16.4.222:34939 172.17.60.201:80 192.168.6.5:80   
TCP 01.23 ESTABLISHED 172.16.4.222:34940 172.17.60.201:80 192.168.6.4:80   
TCP 08.99 ESTABLISHED 172.16.4.222:34941 172.17.60.201:80 192.168.6.5:80

The output shows two connections on the master linux director that are in the TIME_WAIT state, that is they have been closed by the end-user. It also shows one connection in the ESTABLISHED state, that is the end-user and the real-server still have an open connection to each other.

Each of these connections have been synchronised to the slave. Note that on the slave, all the connections are in the ESTABLISHED state. This is due to an optimisation in the LVS code whereby connections are only synchronised when they are in the ESTABLISHED state. This cuts down unnecessary synchronisation overhead as the state of the connections on the slave is not critical.

You can further monitor which linux director is handling connections by adding the following iptables rule to each linux director.

iptables -A INPUT -d 172.17.60.201 -j ACCEPT

This can be monitored using iptables -L INPUT -v -n.

iptables -L INPUT -v -n # On the Active Linux Director
Chain INPUT (policy ACCEPT 1553 packets, 211K bytes)
 pkts bytes target prot opt in  out source    destination         
    5  1551 ACCEPT all  --  *     * 0.0.0.0/0 172.17.60.201

iptables -L INPUT -v -n # On the Stand-By Linux Director
Chain INPUT (policy ACCEPT 2233 packets, 328K bytes)
 pkts bytes target prot opt in  out source    destination         
    0     0 ACCEPT all  --  *     * 0.0.0.0/0 172.17.60.201


To test that connection synchronisation is working correctly open a connection to the virtual service while the master linux director is active. Then cause fail-over to occur, this can be done by a variety of means including powering down the master linux director. At this point the connection should stall. Once the VIP has failed over to the slave linux director the connection should continue.

Streaming is a useful way to test this, as streaming connections by their nature are open for a long time. It also provides intuitive feedback as the video and/or music pause and then continue. It is of note that by increasing the buffer size of the streaming client software the pause can be eliminated.

Problems

The main problem with this implementation is the master/slave relationship. If the master Linux Director fails and then comes back on line, then connections to the slave will not be synchronised to the master. The next time that a fail-over occurs, this will cause cause connections to be terminated. This could be avoided by starting and stopping the master and backup daemons as fail-overs occur. But a peer to peer relationship between the synchronisation daemons would be a cleaner approach.

Connection Synchronisation - New Solution

To improve this situation I have written a new synchronisation daemon for LVS. It works on a peer to peer basis where any node may send or receive synchronisation information.

The new synchronisation daemon runs in user-space rather than the kernel. Information is received from LVS in the kernel via a netlink socket. It is then sent to other nodes using multicast UDP. When a daemon receives information over multicast it reverses this process by sending the information into LVS in the kernel via a netlink socket.

The idea of moving the code to the user-space was to allow more sophisticated synchronisation processing to take place. This is easier to implement and in many ways more appropriately done in user space than the kernel. Given that synchronisation is not a particularly intensive task, there is no particular advantage to keeping it in the kernel.

The code comprises:

Available from www.ultramonkey.org

Running

Installing and compiling is a bit tricky as this is new code and there are a number of support libraries required. Once built, make sure that the LVS kernel synchronisation daemons are not running using ipvsadm -stop-daemon and start the user-space daemon from the command line or using the ip_vs_user_sync_simple init script.

Testing and Debugging

Debugging messages for ip_vs_user_sync_simple are sent to syslog by default and are typically written to /var/log/messages. If the daemon is not functioning correctly, it is recommended to run it with the debug option enabled and have messages logged to the terminal. This can be done my modifying ip_vs_user_sync_simple.conf or on the command line.

ip_vs_user_sync_simple --debug --log_facility -

Testing is as for the existing connection synchronisation code described previously. However, as there is no master/backup relationship connections can be maintained through multiple fail-overs.

Active-Active

Active/Stand-By Linux Directors offer good way to provide high availability. However, if one assumes that a linux director does not fail or get taken down for maintenance very often, then most of the time one linux director will be idle. Arguably this is a waste of resources. It also means that the maximum throughput of the network is limited to that of one linux director.

Having Active-Active linux directors addresses this problem by allowing more than one linux director to load balance connections, for the same virtual services, at the same time.

Figure 9: Active-Active Block Diagram
\includegraphics[width=14cm]{block-diagram}

I have made an implementation of this which works as follows:

Running

Saru can be run directly from the command line or using its own init script. Messages are logged to syslog after startup and these typically appear in /var/log/messages. Saru can log more verbosely by setting the debug option, either in saru.conf or directly on the command line. For debugging purposes this option is recommended in conjunction with having saru log to the terminal.

saru --debug --log_facility -

By default saru waits 30 seconds after startup before joining the cluster. This is to allow time for connection synchronisation to occur when a Linux Director boots up. This can be configured at run time, again either in saru.conf or directly on the command line.

The MAC and IP address of an interface can be set using the ip command.

ip link set eth0 down
ip link set eth0 address 00:50:56:14:03:40
ip link set eth0 up
ip route add default via 172.16.0.254
ip addr add dev eth0 192.168.20.40/24 broadcast 255.255.255.0

Rules to filter out all traffic to the VIP that are not accepted by Saru are inserted using the iptables command. These rules assume that connection synchronisation will be used, If this is not the case then netfilter's connection tracking should be used to ensure that a given connection will always be handled by the same linux director.

iptables -F
iptables -A INPUT -d 172.17.60.201 -p tcp -m saru --id 1 -j ACCEPT
iptables -A INPUT -d 172.17.60.201 -p udp -m saru --id 1 -j ACCEPT
iptables -A INPUT -d 172.17.60.201 -p icmp -m icmp --icmp-type echo-request \
        -m saru --id 1  --sense src-addr -j ACCEPT
iptables -A INPUT -d 172.17.60.201 -p icmp -m icmp --icmp-type ! echo-request \
        -j ACCEPT
iptables -A INPUT -d 172.17.60.201  -j DROP

If LVS-NAT is being used then the following rules are also required to prevent all the Linux Directors sending replies on behalf of the the real servers.

iptables -t nat -A POSTROUTING -s 192.168.6.0/24 -d 192.168.6.0/24 -j ACCEPT
iptables -t nat -A POSTROUTING -s 192.168.6.0/24 -m state --state INVALID \
        -j DROP
iptables -t nat -A POSTROUTING -s 192.168.6.0/24 -m state --state ESTABLISHED \
        -j ACCEPT
iptables -t nat -A POSTROUTING -s 192.168.6.0/24 -m state --state RELATED \
        -j ACCEPT
iptables -t nat -A POSTROUTING -s 192.168.6.0/24 -p tcp -m state --state NEW \
        --tcp-flags SYN,ACK,FIN,RST SYN -m saru --id 1 -j MASQUERADE
iptables -t nat -A POSTROUTING -s 192.168.6.0/24 -p udp -m state --state NEW \
        -m saru --id 1 -j MASQUERADE
iptables -t nat -A POSTROUTING -s 192.168.6.0/24 -p icmp \
        -m icmp --icmp-type echo-request -m state --state NEW \
        -m saru --id 1 --sense dst-addr -j MASQUERADE
iptables -t nat -A POSTROUTING -s 192.168.6.0/24 -p icmp \
        -m icmp --icmp-type ! echo-request -m state --state NEW \
        -j MASQUERADE
iptables -t nat -A POSTROUTING -s 192.168.6.0/24 -j DROP

Testing and Debugging

Which linux director is accepting packets for an individual connection can be monitored using ipvsadm -L INPUT -n -v. The output below shows a connection that was load balanced by Linux Director A.

ipvsadm -L INPUT -n -v # On Linux Director A
Chain INPUT (policy ACCEPT 92541 packets, 14M bytes)
pkts bytes target prot opt in out source    destination         
   5  1551 ACCEPT tcp  --  *  *   0.0.0.0/0 172.17.60.201 saru id 1 sense src-port
   0     0 ACCEPT udp  --  *  *   0.0.0.0/0 172.17.60.201 saru id 1 sense src-port
   0     0 ACCEPT icmp --  *  *   0.0.0.0/0 172.17.60.201 icmp type 8 saru id 1 sense src-addr
   0     0 ACCEPT icmp --  *  *   0.0.0.0/0 172.17.60.201 icmp !type 8 
   0     0 DROP   all  --  *  *   0.0.0.0/0 172.17.60.201

ipvsadm -L INPUT -n -v # On Linux Director B
Chain INPUT (policy ACCEPT 92700 packets, 15M bytes)
pkts bytes target prot opt in out source    destination         
   0     0 ACCEPT tcp  --  *  *   0.0.0.0/0 172.17.60.201 saru id 1 sense src-port
   0     0 ACCEPT udp  --  *  *   0.0.0.0/0 172.17.60.201 saru id 1 sense src-port
   0     0 ACCEPT icmp --  *  *   0.0.0.0/0 172.17.60.201 icmp type 8 saru id 1 sense src-addr
   0     0 ACCEPT icmp --  *  *   0.0.0.0/0 172.17.60.201 icmp !type 8 
   5  1551 DROP   all  --  *  *   0.0.0.0/0 172.17.60.201

Conclusion

LVS is an effective way to implement clustering of Internet services. Tools such as heartbeat, ldirectord and keepalived can be used to give the cluster high availability. There are a number of other techniques that can be used to further enhance LVS clusters including using active feedback to determine the proportion of connections allocated to each of the real servers. As well as connection synchronisation and active active techniques to multiple linux directors to better work together.

LVS itself is a very powerful tool and has many features that were not within the scope of this presentation. These include; firewall marks to group virtual services, specialised scheduling algorithms and various tuning parameters. Beyond that there is much scope for further expanding the functionality of LVS to meet the new needs of users and to reflect the ever increasing complexity of the Internet.

Sample Configuration files for keepalived

Sample configuration file for keepalived master.

global_defs {
   notification_email {
     admin@some.domain
   }
   notification_email_from admin@some.domain
   smtp_server 210.128.90.2
   smtp_connect_timeout 30
   lvs_id LVS_DEVEL
}
vrrp_sync_group VG1 {
    group {
        VI_1
	VI_2
    }
}
vrrp_instance VI_1 {
    state MASTER
    interface eth0
    virtual_router_id 51
    priority 100
    advert_int 1
    authentication {
        auth_type PASS
        auth_pass 1111
    }
    virtual_ipaddress {
        172.17.60.201
    }
}
vrrp_instance VI_2 {
    state MASTER
    interface eth1
    virtual_router_id 52
    priority 100
    advert_int 1
    authentication {
        auth_type PASS
        auth_pass 1111
    }
    virtual_ipaddress {
        192.168.6.1
    }
}
virtual_server 172.17.60.201 80 {
    delay_loop 6
    lb_algo lc
    lb_kind NAT
    nat_mask 255.255.255.0
    !persistence_timeout 50
    protocol TCP
    real_server 192.168.6.4 80 {
        weight 1
        HTTP_GET {
            url {
              path /
              digest 55fd843c4e99e96c1ef28e7dbb10c51b
            }
            connect_timeout 3
            nb_get_retry 3
            delay_before_retry 3
        }
    }
    real_server 192.168.6.5 80 {
        weight 1
        HTTP_GET {
            url {
              path /
              digest 90bfbce6bc089a41f1fddca9aeaba452
            }
            connect_timeout 3
            nb_get_retry 3
            delay_before_retry 3
        }
    }
    sorry_server 127.0.0.1 80
}

Sample Configuration file for keepalived (Slave)

global_defs {
   notification_email {
     admin@some.domain
   }
   notification_email_from admin@some.domain
   smtp_server 210.128.90.2
   smtp_connect_timeout 30
   lvs_id LVS_DEVEL
}
vrrp_sync_group VG1 {
    group {
        VI_1
	VI_2
    }
}
vrrp_instance VI_1 {
    state SLAVE
    interface eth0
    virtual_router_id 51
    priority 100
    advert_int 1
    authentication {
        auth_type PASS
        auth_pass 1111
    }
    virtual_ipaddress {
        172.17.60.201
    }
}
vrrp_instance VI_2 {
    state SLAVE
    interface eth1
    virtual_router_id 52
    priority 100
    advert_int 1
    authentication {
        auth_type PASS
        auth_pass 1111
    }
    virtual_ipaddress {
        192.168.6.1
    }
}
virtual_server 172.17.60.201 80 {
    delay_loop 6
    lb_algo lc
    lb_kind NAT
    nat_mask 255.255.255.0
    !persistence_timeout 50
    protocol TCP
    real_server 192.168.6.4 80 {
        weight 1
        HTTP_GET {
            url {
              path /
              digest 55fd843c4e99e96c1ef28e7dbb10c51b
            }
            connect_timeout 3
            nb_get_retry 3
            delay_before_retry 3
        }
    }
    real_server 192.168.6.5 80 {
        weight 1
        HTTP_GET {
            url {
              path /
              digest 90bfbce6bc089a41f1fddca9aeaba452
            }
            connect_timeout 3
            nb_get_retry 3
            delay_before_retry 3
        }
    }
    sorry_server 127.0.0.1 80
}

Bibliography

1
S. Knight et al.
Rfc 2338: Virtual router redundancy protocol.
http://www.ietf.org/, April 1998.

2
Y. Rekhter et al.
Rfc 1918: Address allocation for private internets.
http://www.ietf.org/, February 1996.

3
Jeremy Kerr.
Using dynamic feeback to optimise load balancing decisions.
http://www.redfishsoftware.com.au/projects/feedbackd/lca-paper.pdf, January 2003.

4
Netfilter Core Team.
Netfilter - firewalling, nat and packet mangling for linux 2.4.
http://www.netfilter.org/, 2003.



Horms 2004-06-23