Speaker
Mr
Vladimir Bahyl
(CERN IT-FIO)
Description
Availability approaching 100% and response time converging to 0 are two factors that
users expect of any system they interact with. Even if the real importance of these
factors is a function of the size and nature of the project, todays users are rarely
tolerant of performance issues with system of any size.
Commercial solutions for load balancing and failover are plentiful. Citrix NetScaler,
Foundry ServerIron series, Coyote Point Systems Equalizer and Cisco Catalyst SLB
switches, to name just a few, all offer industry standard approaches to these
problems. Their solutions are optimized for standard protocol services such as HTTP,
FTP or SSH but it remains difficult to extend them for other kinds of application. In
addition to this, the granularity of their failover mechanisms are per node and not
per application daemon, as is often required. Moreover, the pricing of these devices
for small projects is uneconomical.
This paper describes the design and implementation of the DNS load balancing and
failover mechanism currently used at CERN. Our system is based around SNMP, which is
used as the transport layer for state information about the server nodes. A central
decision making service collates this information and selects the best candidate(s)
for the service. IP addresses of the chosen nodes are updated in DNS using the DynDNS
mechanism.
The load balancing feature of our system is used for variety of standard protocols
(including HTTP, SSH, (Grid)FTP, SRM) while the (easily extendable) failover
mechanism adds support for applications like CVS and databases. The scale, in terms
of the number of nodes, of the supported services ranges from a couple (2-4), up to
around 100. The best known services using this mechanism at CERN are LXPLUS and
CASTORGRID.
This paper also explains the advantages and disadvantages of our system, and advice
is given about when it is appropriate to be used.
Last, but not least, given the fact that all components of our system are build
around freely available open source products, our solution should be especially
interesting in low resource locations.
Primary author
Mr
Vladimir Bahyl
(CERN IT-FIO)
Co-author
Mr
Nicholas Garfield
(CERN IT-CS)
