Modernization of Classic Structure Supercomputers
If the high performance computing complex is already used but its resources are insufficient to solve current tasks there is a need to update the system. This is a very fine issue as a rule connected to a range of difficulties.
Here conventional extensive approach to growing of computing capacities is possible.
As a rule, systems with maximum possible configuration of processors and memory are used as supercomputer nodes. That is why updating includes only increase of number of computing nodes (horizontal scaling). Often the nodes of the same type as existing ones are unavailable as of the moment of update. And new nodes, as a rule, are constructed on the basis of new architecture, state-of-the-art components, they use faster memory, input/output interfaces. This leads to heterogeneity undesirable for unified supercomputer. Also it is necessary to consider that growth of complex performance with extension of node quantity is not linear and the effect of node addition will be lesser and lesser with every additional node. At the same time performance buildup by horizontal scaling leads to significant growth of energy consumption which is not always permissible.
The most optimal way to update the supercomputer is reorganization by adjustment to hybrid infrastructure.
At first it seems not very realizable as far as computing nodes were initially not intended for graphic processors. But Open Technologies has a solution. To implement this possibility it is just necessary for the nodes of existing cluster to allow installing special interface card of PCI-E standard. This card being installed into cluster nodes allows connecting graphic processing units (GPU) externally. Separate housing sized as standard 1U server comprises up to 4 standard GPUs and up to 16 GPUs in 3U for high density construct. At this they have independent power connection, deducted cooling system and environmental controls.
Hybrid supercomputer constructed using independent constructs for CPU and GPU parts is not only an elegant solution of supercomputer updating. It has a range of unique characteristics. Independent life cycles of central and graphic processing units allow flexibility of maintenance and updating, optimal distribution of heating sources in racks thus contributing to more advantageous component operation modes and finally having positive effect for their life cycle.
That is why approach of GPU separation into independent construct is used by specialists of Open Technologies not only to ensure updating of the existing systems but also when constructing standard configurations described in preceding section.
Application of specialized constructs for deployment of graphic processing units allows maximum transparency of transition from supercomputer of classic scheme to hybrid supercomputer preserving all its advantages. This is a mechanism to protect customer's investments and reduce significantly the aggregate cost of possession of high performance computing hardware.