HA-OSCAR: the Birth of Highly Available OSCAR

One important factor to consider is the time it takes to build, boot and have the cluster ready to service requests. This is not a major issue for small clusters, but as we move to large installations of 256 nodes and higher, having the capabilities of installing and booting all cluster nodes in an automated and timely manner becomes an asset. HA-OSCAR is considering implementing hierarchical clustering by dividing the cluster into multiple zones. This type of experimentation also can be helpful in identifying the slow processes in the system installation procedure, which allows us to bring it up to speed. LinuxBIOS, for instance, can be included in place of the normal BIOS—with a little bit of hardware initialization and a compressed Linux kernel that can be booted from a cold start—to achieve faster startup times. The upcoming OSCAR release uses multicast technology, which was tested on about 500 nodes, to speed up install times and return impressive numbers. HA-OSCAR plans to adopt this method as the base install mechanism and improve on it.

Similar to the base OSCAR installation, users of HA-OSCAR have the freedom of deciding which application packages to install. By default, HA-OSCAR automatically installs the essential parts to build a cluster and then prompts the user to select the applications they want.

The installation procedure takes into consideration any existing configuration and the packages already installed on the node. Some packages are sensitive to certain system libraries, such as glibc. Users should be aware that installing HA-OSCAR may require them to upgrade their systems based on such dependencies. In the same manner, a de-installation procedure is provided to clean up every HA-OSCAR-specific addition without disturbing the system integrity. This option is important for users who want to test only HA-OSCAR.

It is also worth mentioning that package install and uninstall options are available in the base OSCAR release since v2.0, and a newly enhanced version is coming out soon.

HA-OSCAR plans to investigate the possibility of providing mechanisms for selective network software upgrades without bringing down the system. Network upgrades are an interesting way of patching an operating system and its applications. As an example, most Linux distributions now come with an automatic network upgrade that eases this tedious administrative task. In the case of administrating a large cluster, HA-OSCAR users can use such a feature to upgrade their application version seamlessly, without service interruption. Network upgrade simplifies cluster administration and promotes better software management across all computing nodes.

In addition, HA-OSCAR provides a tool that allows users to change the configuration of the cluster at runtime by using a tool somewhat similar to LinuxConf. This is still a basic idea that will be investigated further in the near future.

Generally, one cannot trust a computing system if there is no backup or recovery mechanism. For mission-critical applications, including telecom applications, it is important to be able to recover from any software or hardware failure. Thus, providing efficient backup and recovery mechanisms is an essential part of any HA system.

In case a disaster occurs, recovery ability and speed are critical. Every time HA-OSCAR is completely re-installed or the kernel updated, ghost images of before and after are saved in a designated location on a backup server and tape. Ghost for Unix takes a snapshot of an old and new kernel, gzips it and sends the image to the secondary head node as well as to a predefined disaster recovery site. Important data as well as application and configuration files also can be included in the ghost image. Normally, tape backup schedules include nightly snapshots for incremental images and weekly snapshots for full images. For faster recovery and highly reliable backups, ghost imaging, file journaling and data replication are implemented.

One goal of HA-OSCAR is to be deployed optionally as a Web server cluster providing highly available Web services to a large number of clients. One step toward this goal is to set up a Web server, such as Apache, on every node; Apache can be one of the packages copied to the nodes. Then, a single IP interface is provided for the cluster, possibly using LVS Direct Routing, because it has proven to be the scalable implementation.