Wednesday 24 August 2011

GPU Clusters

GPU Clusters

Improving cluster performance is all about using the right hardware. Administrators could implement improvements in their clusters with the right equipments that are geared towards improvement of computing. However, the changes required to improve cluster performance could be costly for the business.

Additional powerful nodes, better infrastructure and connections and updated application to control the nodes are only some of the requirements in improving the computing performance of the clusters. Some businesses are just forced to contend with what they have and wait for additional revenues before they can implement certain changes.

But there is actually an option for administrators to effectively improve the computing performance of their clusters without having to spend too much on upgrading. Instead of adding nodes that could be costly for businesses, the nodes that are currently used are upgraded to improve their computing performance.

The challenge in improving the nodes is that the slots for RAM which provides an upgraded processing power for the nodes are already filled up. A single node could have as much as 8GB of RAM but it might not be enough to cater to large number of data and function requests.

Instead of upgrading RAM which is almost impossible for now; administrators have looked into updating the GPU or the Graphics Processing Unit of the node. Although GPU is used primarily for improving videos, they can be harnessed to assist computing in the nodes.

Forms of GPU Clusters

GPU powered clusters could be differentiated based on the types of GPU installed in the nodes.

The first form of GPU powered cluster is the heterogeneous type of GPU clustering. This refers to the installation of different types of GPUs installed in a node. They could come from different manufacturer and might even have different processing power. The advantage of this form of GPU clustering is the cost efficiency since a powerful GPU could be assisted by less powerful GPU.

The next form of GPU powered cluster is the homogeneous form. As the name suggests, the GPUs installed in the node are uniform in processing power, make and model. Although a little bit expensive for powerful GPUs, they provide the extra processing power every developer expects.

Careful consideration should be done when working with GPUs. Most GPUs have hardware requirements for them to be fully utilized. This is especially true when dealing with heterogeneous GPU powered clusters as each might have different hardware requirements.

Connection Requirements

For the GPU to properly work as performance boosters when installed, additional connection is required. This is not only to optimize the performance of each nodes but the additional installation will mean powerful interaction among nodes.

The good news is that there are no additional software requirements just to ensure that the optimization of a node through GPU could be achieved. Connection will not also require additional software although control of nodes with GPU will have a special application which will be outlined later.

An additional Ethernet switch or a communications link between GPUs is required. The communications link is more ideal as they provide additional control for administrators.

Operational Requirements

To properly implement GPU powered clustering, the following software should be present:

1. The Right Operating System – Although the operating system could never influence the processing speed of GPUs, they are installed to ensure compatibility of the GPU with the rest of the node.

2. Drivers – GPUs will never work without its drivers properly installed. They are provided by the manufacturers and should be present in each node to ensure performance even though the clusters have homogeneous GPUs.

3. API – This is the unique requirement for administrators before they could fully implement GPU as additional source of processing power. There are APIs today that are geared towards proper communication of GPUs within the cluster. An API could be considered as additional software that could be implemented together with the application that controls the general environment of the cluster.

4. Mapping – A unique behavior of GPU as additional source of power is that they can’t be easily controlled independently through an application. However, mapping could be implemented wherein algorithms are added to suggest behavior of the GPU when faced with a specific situation.

GPU powered clustering is an efficient but cost effective upgrade that administrators could implement in the nodes. With the right applications and few hardware additions, their power could be harnessed to improve computing speed.

No comments:

Post a Comment

VMware Cloud Learning Video's

Here is a nice summary list of all VMworld US 2018 Breakout session with the respective video playback & download URLs. Enjoy! Bra...