This dissertation discusses the configuration of the Gollach cluster. The system has a shared disk spread over all machines, and uses MOSIX to share load. All machines boot off a single server.
This thesis describes the Gollach cluster. The Gollach is an eight machine computing cluster that is aimed at being a general purpose computing resource for research purposes. This includes image processing and simulations. The main quest in this project is to create a cluster server that gives increased computational power ad a unified system image (at several levels) without requiring the users to learn specialised tricks. At the same time the cluster must not be tasking to administer.
The making of this cluster aims to give a single image on several levels of the computer, which are: processes, I/O (Input/Output) and storage. To achieve this the cluster computer is built together on a unified backbone that ensures sharing on the levels mentioned above. It makes use of diskless booting, NFS-root (Network File System), MOSIX and LVS (Linux Virtual Server) to achieve this. MOSIX migrates processes and balances load. LVS distributes users on the cluster server. Diskless booting and NFS-root file system ensure that all the machines on the cluster share a single disk image. Mass storage has been aggregated with the help of NBD (Network Block Device) and software RAID (Redundant Array of Independent Disks).
The Gollach has been built and it provides a general computing resourec for multiple users. Administration turns out to be at almost the same level as a single Linux server due to the use of ClusterNFS. ClusterNFS is an NFS server which provides name translation. Therefore machine specific files can be kept unique on the file server. User distribution using LVS has worked as expected, users’ requests are always attached to the machine with the least connections. The mass storage device has been tested to work with RAID-0. The cluster has been made to run conventional programs as a conventional server, here it works to expectation with no visible difference to a single machine server. Parallelism with the help of MOSIX has been observed and non-I/O bound processes have experienced a speed-up due to the process migration. Disk I/O processes suffer though, from the speed descaling found with multiple simultaneous reads/writes.