Load balancer scheduler algorithm

At present, there are 3 load balancer scheduler algorithms available for use: Request Counting, Weighted Traffic Counting and Pending Request Counting. These are controlled via the lbmethod value of the Balancer definition. See the ProxyPass directive for more information.

Load balancer stickyness

The balancer supports stickyness. When a request is proxied to some back-end, then all following requests from the same user should be proxied to the same back-end. Many load balancers implement this feature via a table that maps client IP addresses to back-ends. This approach is transparent to clients and back-ends, but suffers from some problems: unequal load distribution if clients are themselves hidden behind proxies, stickyness errors when a client uses a dynamic IP address that changes during a session and loss of stickyness, if the mapping table overflows.

The module mod_proxy_balancer implements stickyness on top of two alternative means: cookies and URL encoding. Providing the cookie can be either done by the back-end or by the Apache web server itself. The URL encoding is usually done on the back-end.

Examples of a balancer configuration

Before we dive into the technical details, here's an example of how you might use mod_proxy_balancer to provide load balancing between two back-end servers:

Another example of how to provide load balancing with stickyness using mod_headers, even if the back-end server does not set a suitable session cookie:

Request Counting Algorithm

Enabled via lbmethod=byrequests, the idea behind this scheduler is that we distribute the requests among the various workers to ensure that each gets their configured share of the number of requests. It works as follows:

lbfactor is how much we expect this worker to work, or the workers's work quota. This is a normalized value representing their "share" of the amount of work to be done.

lbstatus is how urgent this worker has to work to fulfill its quota of work.

The worker is a member of the load balancer, usually a remote host serving one of the supported protocols.

We distribute each worker's work quota to the worker, and then look which of them needs to work most urgently (biggest lbstatus). This worker is then selected for work, and its lbstatus reduced by the total work quota we distributed to all workers. Thus the sum of all lbstatus does not change(*) and we distribute the requests as desired.

If some workers are disabled, the others will still be scheduled correctly.

for each worker in workers
    worker lbstatus += worker lbfactor
    total factor    += worker lbfactor
    if worker lbstatus > candidate lbstatus
        candidate = worker

candidate lbstatus -= total factor

If a balancer is configured as follows:

And b gets disabled, the following schedule is produced:

That is it schedules: a c d a c d a c d ... Please note that:

Has the exact same behavior as:

This is because all values of lbfactor are normalized with respect to the others. For:

worker b will, on average, get 4 times the requests that a and c will.

The following asymmetric configuration works as one would expect:

That is after 10 schedules, the schedule repeats and 7 a are selected with 3 b interspersed.

Weighted Traffic Counting Algorithm

Enabled via lbmethod=bytraffic, the idea behind this scheduler is very similar to the Request Counting method, with the following changes:

lbfactor is how much traffic, in bytes, we want this worker to handle. This is also a normalized value representing their "share" of the amount of work to be done, but instead of simply counting the number of requests, we take into account the amount of traffic this worker has seen.

If a balancer is configured as follows:

Then we mean that we want b to process twice the amount of bytes than a or c should. It does not necessarily mean that b would handle twice as many requests, but it would process twice the I/O. Thus, the size of the request and response are applied to the weighting and selection algorithm.

Pending Request Counting Algorithm

Enabled via lbmethod=bybusyness, this scheduler keeps track of how many requests each worker is assigned at present. A new request is automatically assigned to the worker with the lowest number of active requests. This is useful in the case of workers that queue incoming requests independently of Apache, to ensure that queue length stays even and a request is always given to the worker most likely to service it fastest.

In the case of multiple least-busy workers, the statistics (and weightings) used by the Request Counting method are used to break the tie. Over time, the distribution of work will come to resemble that characteristic of byrequests.

This algorithm is available in Apache HTTP Server 2.2.10 and later.

Exported Environment Variables

At present there are 6 environment variables exported:

BALANCER_SESSION_STICKY

This is assigned the stickysession value used for the current request. It is the name of the cookie or request parameter used for sticky sessions

BALANCER_SESSION_ROUTE

This is assigned the route parsed from the current request.

BALANCER_NAME

This is assigned the name of the balancer used for the current request. The value is something like balancer://foo.

BALANCER_WORKER_NAME

This is assigned the name of the worker used for the current request. The value is something like http://hostA:1234.

BALANCER_WORKER_ROUTE

This is assigned the route of the worker that will be used for the current request.

BALANCER_ROUTE_CHANGED

This is set to 1 if the session route does not match the worker route (BALANCER_SESSION_ROUTE != BALANCER_WORKER_ROUTE) or the session does not yet have an established route. This can be used to determine when/if the client needs to be sent an updated route when sticky sessions are used.

Enabling Balancer Manager Support

This module requires the service of mod_status. Balancer manager enables dynamic update of balancer members. You can use balancer manager to change the balance factor of a particular member, or put it in the off line mode.

Thus, in order to get the ability of load balancer management, mod_status and mod_proxy_balancer have to be present in the server.

To enable load balancer management for browsers from the example.com domain add this code to your httpd.conf configuration file

You can now access load balancer manager by using a Web browser to access the page http://your.server.name/balancer-manager

Details on load balancer stickyness

When using cookie based stickyness, you need to configure the name of the cookie that contains the information about which back-end to use. This is done via the stickysession attribute added to either ProxyPass or ProxySet. The name of the cookie is case-sensitive. The balancer extracts the value of the cookie and looks for a member worker with route equal to that value. The route must also be set in either ProxyPass or ProxySet. The cookie can either be set by the back-end, or as shown in the above example by the Apache web server itself.

Some back-ends use a slightly different form of stickyness cookie, for instance Apache Tomcat. Tomcat adds the name of the Tomcat instance to the end of its session id cookie, separated with a dot (.) from the session id. Thus if the Apache web server finds a dot in the value of the stickyness cookie, it only uses the part behind the dot to search for the route. In order to let Tomcat know about its instance name, you need to set the attribute jvmRoute inside the Tomcat configuration file conf/server.xml to the value of the route of the worker that connects to the respective Tomcat. The name of the session cookie used by Tomcat (and more generally by Java web applications based on servlets) is JSESSIONID (upper case) but can be configured to something else.

The second way of implementing stickyness is URL encoding. The web server searches for a query parameter in the URL of the request. The name of the parameter is specified again using stickysession. The value of the parameter is used to lookup a member worker with route equal to that value. Since it is not easy to extract and manipulate all URL links contained in responses, generally the work of adding the parameters to each link is done by the back-end generating the content. In some cases it might be feasible doing this via the web server using mod_substitute. This can have negative impact on performance though.

The Java standards implement URL encoding slightly different. They use a path info appended to the URL using a semicolon (;) as the separator and add the session id behind. As in the cookie case, Apache Tomcat can include the configured jvmRoute in this path info. To let Apache find this sort of path info, you need to set scolonpathdelim to On in ProxyPass or ProxySet.

Finally you can support cookies and URL encoding at the same time, by configuring the name of the cookie and the name of the URL parameter separated by a vertical bar (|) as in the following example:

If the cookie and the request parameter both provide routing information for the same request, the information from the request parameter is used.

Troubleshooting load balancer stickyness

If you experience stickyness errors, e.g. users lose their application sessions and need to login again, you first want to check whether this is because the back-ends are sometimes unavailable or whether your configuration is wrong. To find out about possible stability problems with the back-ends, check your Apache error log for proxy error messages.

To verify your configuration, first check, whether the stickyness is based on a cookie or on URL encoding. Next step would be logging the appropriate data in the access log by using an enhanced LogFormat. The following fields are useful:

%{MYCOOKIE}C

The value contained in the cookie with name MYCOOKIE. The name should be the same given in the stickysession attribute.

%{Set-Cookie}o

This logs any cookie set by the back-end. You can track, whether the back-end sets the session cookie you expect, and to which value it is set.

%{BALANCER_SESSION_STICKY}e

The name of the cookie or request parameter used to lookup the routing information.

%{BALANCER_SESSION_ROUTE}e

The route information found in the request.

%{BALANCER_WORKER_ROUTE}e

The route of the worker chosen.

%{BALANCER_ROUTE_CHANGED}e

Set to 1 if the route in the request is different from the route of the worker, i.e. the request couldn't be handled sticky.

Common reasons for loss of session are session timeouts, which are usually configurable on the back-end server.

The balancer also logs detailed information about handling stickyness to the error log, if the log level is set to debug or higher. This is an easy way to troubleshoot stickyness problems, but the log volume might be to high for production servers under high load.