Construction of Hybrid Supercomputers
Historically the term "High Performance Computing" (supercomputer) was connected to something very expensive, energy-intensive, occupying a lot of space. Development of the industry leads to changes of approach to construction of supercomputers. They become efficient in terms of energy, space occupied in technological premise and financially.
Times when only very large organizations could afford to have high performance computing are passing. First of all, this is, of course, connected to the fact that the supercomputer became cheaper and more effective. In other words, capital and operational expenses of the system possession reduced significantly. Growing popularity of parallel technologies leads to situation when more and more manufacturers adapt their software to work with platforms of such architecture.
These last years the growth of interest to supercomputers has been connected to emergence of so called hybrid supercomputers at the market. These are supercomputers which use specialized processors, particularly, graphic ones, along with conventional architecture of central processing units for calculations. Open Technologies proposes services on creation of hybrid architecture supercomputers on the basis of graphic processors which are noted for lower energy consumption, lower cost and compactness as compared to conventional ones.
In these computers the tasks are solved using processors initially developed to calculate graphics. The thing is about the number of simultaneous operations processed. The number of kernels of state-of-the-art graphic processor is 2 times more than number of kernels of conventionally constructed processor. No doubt, these kernels are able to implement limited number of functions and have limited core memory as compared to central processor and its kernels. But this is enough and to spare to solve huge number of computations allowing paralleling of calculation algorithms of the resources and the number of kernels finally allows achieving tens and hundreds time enhancement in performance.
For example, earlier solution of seismic analysis tasks required supercomputer consisting of 2,000 computing nodes consuming 1,200 kW estimated to about $8 mln. But state-of-the-art hybrid supercomputer able to solve the same task requires 30 times less space, is almost 30 times energy efficient and 20 times cheaper.
However, the supercomputer is not only processors, memory and components of interaction between nodes. Unification of computers into single system working on one task is not a usual process; it requires using specialized hardware and software solutions. For example, exchange of messages between processes in both hybrid and conventional systems MPI protocols are used. Choice of certain protocol construction depends on task. Every certain task requires optimum these or that MPI construction. The chosen approach to organization of internode interaction channels has significant influence on the total performance and efficiency of computing resource use. In some cases usual Ethernet is enough but sometimes it is necessary to construct special low latency network, for example, based on Infiniband or less common specialized communication media like Quadrix and Myrinet. High efficiency of supercomputer operation can be reached only by customizing of architecture to customer's tasks. Our specialists got used to work just like this.
The supercomputer task software can be both commercial software and customized program packages developed to solve certain task in optimal way.
The software shall certainly consider the supercomputer hybrid architecture and use its resources. Almost all leading manufacturers of commercial software work in this field. The software which can work with hybrid architecture includes widespread mathematic packages like MathWorks MATLAB, Wolfram Mathematica; design automation packages like ANSYS, Abaqus; 3D modeling, animation, rendering packages like 3Ds Max, Maya; graphic packages like Photoshop, After Effects, Premiere; molecular dynamics computation packages like Amber, NAMD, etc.
Specialized software developed for classic high performance computing requires adaptation when transmitting it to hybrid architecture but this process is not as difficult as it seems to be. The general idea of adaptation is separation of code parts which can be implemented in parallel at specialized computer and provision of correct sequence so that the graphic process resources are used in optimal way. The primary assessment of transfer of software to hybrid supercomputer can be received with packages automating adaptation of software, such as, for example, CAPS. Later the effect of transfer of tasks implemented to hybrid architecture can be strengthened by modifications of programs made by developers.
To accelerate submersion of developer teams into the world of supercomputers and simplify hybrid technologies introduction process Open Technologies provides trainings for customers' programmers.
With hybrid supercomputer you receive:
- Use of hybrid supercomputers allows you to reach performance rated at tens and hundreds of teraflops and ensure new level of task solving.
- Energy consumption and space occupied by the supercomputer are within due limits.
- Cost of hybrid supercomputer of primary level is almost the same as cost of several professional working stations.