Products


 Contents



 Designing for Maximum Performance


ClusterVision's cluster design philosophy is to aim for maximum performance, stability and quality while keeping prices low. This goal is achieved by carefully selecting hardware components of the best quality and combining this with the best Linux-based or Windows-based cluster software on the market: the ClusterVisionOS™ or Windows CCS.

The performance of a cluster is affected by almost every individual hardware component, but most notably by the CPU, motherboard chipset, memory and network interconnect. selecting the right combination of components is crucial to achieve an optimally performing cluster. Choosing wrong combinations of components can seriously limit the performance of your cluster. This happens, for example, when a motherboard chipset is chosen that limits the PCI bandwidth available to a high-speed network interconnect.

The performance of a cluster is of course also affected by the operating system. We optimise the ClusterVisionOS™ for maximum performance running typical scientific and engineering applications. Optimisations may be applied to the Linux kernel, motherboard BIOS, hardware firmware, drivers, libraries and compilation options.


 Designing for Your Applications


The best combination of hardware components always depends on how your particular application code works:
  • Does it perform more floating point or more integer operations.
  • Does it use large amounts of data and it what way.
  • Does it access this data in a predictable or a random fashion.
  • Can the code easily be parallellised.
  • Is it more efficient to use a fine- or a coarse grained parallelism, etc.
If you use well-known scientific application codes, the chances are one of our in-house or consultant application specialists have previous experience running those codes on a cluster. We can then advise you on the best cluster design for your applications.

If we are not familiar with your particular application code, but you can tell us how your code works, how you read your input data, which computational algorithms are used, etc., then we can often predict reasonably well what combination of hardware components will work best.

Often we advise to first benchmark your application code on a real cluster. We can then provide you with an account on one of our test clusters, or we can benchmark your code for you.



Why Not to Do it Yourself
Are you considering to build your own cluster? Give us a call or drop us an email. We are happy to give you advice and we may be able to provide you with the right hardware at the right price. However, there are many advantages to buying a turnkey cluster solution:

1) Assembling, configuring and fine-tuning a cluster which is easy to use and administrate, and which is stable and secure, requires considerable Linux expertise and costs time, considerable time.

2) When something breaks, when your application doesn't work with your Linux installation, or when your master node breaks down and you have no backup, you're on your own.

3) Without having access to a wide range of hardware components and without having direct contacts with hardware manufacturers, it can be difficult to exclude any hardware incompatibilities between your components.

4) If you order individual hardware components to assemble a cluster yourself, you are bound to receive some faulty components. Diagnosing the problems caused by faulty components can be time-consuming and frustrating. (Read more about our quality and testing procedures.)
 Designing for Minimum Space


We offer various server solutions that minimise use of rack space, including twin servers, blade servers and half-depth servers.

Twin Servers
Twin servers have two complete and independent servers in one 1U chassis. Each server can have a single- or dual-socket motherboard. The servers usually only share the chassis and the power supply. The big advantages of twin servers are the space saving (half the space of a regular 1U server), and the low power consumption. The power consumption is lower than regular 1U servers because two servers share the same power supply, which is more efficient. Twin servers are available with Intel and AMD CPUs and can contain InfiniBand on the motherboard.

Half-Depth Servers
Half-depth servers are very small 1U servers with a regular 19" width, but a short depth. The short depth allows the servers to be placed in a short-depth 19" rack, or depending on the design, in a back-to-back arrangement. Half-depth servers are available with Intel and AMD CPUs.

Blade Servers
Blade servers consist of a large chassis with multiple small blades placed vertically in the chassis. The blades usually share one or more redundant power supplies and fans. Often the chassis also contains Ethernet switches or Ethernet pass-through modules and IPMI compatible management modules. Blades are available with Intel and AMD CPUs. InfiniBand and Myri-10G solutions can be integrated in most blade solutions.



  Twin servers

 Half-depth server

   Blade servers
 Power and Air-conditioning Requirements


When we design your cluster, we take any power and air-conditioning limitations you may have into account to try to come with a design which meets those requirements. As early as possible in the design process, we will provide you with estimates of the power and air-conditioning requirements for your cluster.

A cluster installation is often preceded by a site survey by one of our qualified engineers. The engineer will discuss possible installation locations, delivery access, and power and air-conditioning requirements. If additional power or air-conditioning is required, we can implement this through one of our specialist power or air-conditioning sub-contractors.

We offer a choice between air-cooled and water-cooled rack solutions. Water-cooled solutions are most energy-efficient but more costly for their initial purchase price. From a long-term TCO perspective water-cooled racks are generally more cost-effective than air-cooled racks.



Water-cooled rack
 Designing for Redundancy & High Availability


ClusterVision offers fully redundant cluster designs for 99.99% guaranteed 24/7 uptime. In our redundant cluster designs, every potential point of failure is replaced by some form of hardware redundancy.

Due to the way the ClusterVisionOS™ is designed, cluster operation is not seriously affected by one ore more slaves going off-line. If a slave node goes off-line, the ClusterVisionOS will mark the node as off-line and exclude it from the cluster.

Some clusters are delivered with spare nodes which are kept with the cluster to be swapped in case of a node failure. Both master and slave spare nodes can be provided.

We design most of our master nodes with redundant power supplies and double hot or cold swap hard disks configured as a RAID level 1 mirror. In this RAID configuration, all data is simultaneously stored on both disks and the master node continues operating if a disk fails.

We offer two types of failover systems for master nodes; an active and a passive failover system.

Active/Passive Failover System
The active/passive failover system uses a second master node which constantly monitors the activity of the primary master node. In the active failover system, if the primary master node fails, the secondary master node takes over immediately and without intervention. In the passive failover system, if the primary master node fails, the secondary master node only takes over after intervention by the system administrator.Depending on the usage of the cluster, the required level of control and preferences of the system administrator, the passive or active system may be most appropriate.



Cluster with an active/passive
failover system
Copyright 2008 ClusterVision BV | Legal Notice |