Latest trends in Cluster Computing

Cluster computing has gained attention for a long time now. It is one of the most powerful solutions to tackle the ravenous computational power required by grand challenge applications. Contemporary supercomputing processing nodes employed for clustered architecture noticeably has yielded very large cluster size. Consequently, the cluster architecture designed has also been fairly complex. Therefore a constructive change in the design of such processing nodes can help devise superior cluster architecture capable of enhanced application-level computations.

The internal architecture of processing nodes (of conventional supercomputing cluster) is generally homogeneous and ALU based as in the Blue Gene/L (PowerPC 440) and NASA Columbia (Itanium). Even in chip multiprocessors the cores are replicated processors with common cache shared through non-uniform memory access. Such designs definitely increase the processing power of the node as such, but in the milieu of mapping different types of applications on the cluster, computation might prove to be rather complex, in that the complexity is attributed to the depth of basic operations handled by the cores and hence by the processing nodes. When the basic operations handled by the node are at a much higher level, computational demand can be met with greater efficiency resulting in relatively reduced complexity of the cluster architecture. In this context, heterogeneous multi-core node architecture is envisioned.

The traditional methods of designing high performance node architectures meant for supercomputing clusters, do not deal with design flexibility. The node architecture design has been ad hoc rather than holistic, in that a particular design has revolved around a distinct architectural feature (eg. PIM technology, or scalar processing, or vector processing or special purpose ASIC as in MDGRAPE and so on). This would probably curtail some other features that could otherwise compliment the node architecture under design.