Author Topic: Fiber Optics Based Parallel Computer Architecture  (Read 4376 times)

0 Members and 1 Guest are viewing this topic.


  • International Journal of Scientific and Engineering Research
  • Administrator
  • Jr. Member
  • *****
  • Posts: 89
  • Karma: +0/-1
  • Research Paper Publishing
    • View Profile
    • International Journal of Scientific and Engineering Research
Fiber Optics Based Parallel Computer Architecture
« on: January 22, 2011, 06:18:17 am »
International Journal of Scientific & Engineering Research, Volume 1, Issue 2, November-2010
ISSN 2229-5518

Computer systems that use optical fiber, in particular the Parallel Sysplex architecture from IBM. Other applications do not currently use optical fiber, but they are presented as candidates for optical interconnect in the near future, such as the Power- Parallel supercomputers which are part of the Advanced Strategic Computing Initiative (ASCI). Many of the current applications for fiber optics in this area use serial optical links to share data between processors, although this is by no means the only option. Other schemes including plastic optics, optical backplanes, and free space optical interconnects Towards the end of the paper, we also provide some speculation concerning machines that have not yet been designed or built but which serve to illustrate potential future applications of optical interconnects. Because this is a rapidly developing area, we will frequently cite Internet references where the latest specifications and descriptions of various parallel computers may be found.

Computer engineering often presents developers with a choice between designing a computational device with a single powerful processor (with additional special-purpose coprocessors) or designing a parallel processor device with the computation split among multiple processors that may be cheaper and slower. There are several reasons why a designer may choose a parallel architecture over the simpler single processor design. Before each reason, and other categorizing methods in this paper we will have a letter code, A, which we will use to categorize architectures we describe in other sections of the paper.

1. Speed - There are engineering limits to how fast any single processor can compute using current technology. Parallel architectures can exceed these limits by splitting up the computation among multiple processors.
2. Price - It may be possible but prohibitively expensive to design or purchase a single processor machine to perform a task. Often a parallel processor can be constructed out of off-the-shelf components with sufficient capacity to perform a computing task.
3. Reliability - Multiple processors means that a failure of a processor does not prevent computation from continuing. The load from the failed processor can be redistributed among the remaining ones. If the processors are distributed among multiple sites, then even catastrophic failure at one site (due to natural or man-made disaster, for example) would not prevent computation from continuing.
4. Bandwidth - Multiple processors means that more bus bandwidth can be processed by having each processor simultaneously use parts of the bus bandwidth.
5. Other - Designers may have other reasons for adding parallel Processing not covered above.
Current parallel processor designs were motivated by one or more of these needs. For example, the parallel Sysplex family was motivated by reliability and speed, the Cray XMP was primarily motivated by speed, the BBN butterfly was designed with bandwidth considerations in mind, and the transputer family was motivated by price and speed. After a designer has chosen to use multiple processors he must make several other choices like processors. Number of processors, network topology
The product of the speed of the processors and the number of processors is the maximal processing power of the machine (for the most part unachievable in real life). The effect of network topology is subtler.

Network topologies control communication between machines. While most multiprocessors are connected with ordinary copper-wired buses, we believe that fiber optics will be the bus technology of the future. Topology controls how many computers may be necessary to relay a message from one processor to another. A poor network topology can result in bottlenecks where all the computation is waiting for messages to pass through a few very important machines. Also, a bottleneck can result in unreliability with failure of one or few processors causing either failure or poor performance of the entire system.
Four kinds of topologies have been popular for multiprocessors. They are
  Full connectivity using a crossbar or bus. The historic C.mmp processor used a crossbar to connect the processors to memory (which allowed them to communicate). Computers with small numbers of processors (like a typical parallel Sysplex system or tandem system) can use this topology but it becomes cumbersome with large (more than 16) processors because every processor must be able to simultaneously directly communicate with every other. This topology requires a fan in and fan out proportional to the number of processors, making large networks difficult.
Pipeline where the processors are linked together in a line and information primarily passes in one direction. The CMU Warp processor was a pipelined multiprocessor and many of the first historical multiprocessors, the vector processors, were pipelined multiprocessors. The simplicity of the connections and the many numerical algorithms that are easily pipelined encourage people to design these multiprocessors. This topology requires a constant fan in and fan out, making it easy to lay out large numbers of processors and add new ones.
Torus and Allied topologies where an N processor machine requires √N processors to relay messages. The Goodyear MPP machine was laid out as a torus. Such topologies are easy to layout on silicon so multiple processors can be placed on a single chip and many such chips can be easily placed on a board. Such technology may be particularly appropriate for computations that are spatially organized. This topology also has constant fan in and fan out. Adding new processors is not as easy as in pipelined processors but laying out this topology is relatively easy. Because of the ease of layout sometimes this layout is used on chips and then the chips are connected in a hypercube.
Hypercube and Butterfly topologies have several nice properties that have lead to their dominating large-scale multiprocessor designs. They are symmetric so no processor is required to relay more messages than any other is. Every message need only be relayed through log (N) processors in an N processor machine and messages have multiple alternate routes, increasing reliability under processor failure and improving message routing and throughput. Transputer systems and the BBN butterfly were some of the first multiprocessors that adapted this type of topology. This topology has a logarithmic fan out and that can complicate layout when the size of the processor may grow over time. There is an alternative topology called cube-connected cycles that has the same efficient message passing properties as the hypercube topology but constant fan out, easing layout considerably.
Exotic - There are a variety of less popular but still important topologies one can use on their network.
The more efficient and fast the bus technology is, the simpler the topology can be. A really fast bus can simply connect all the processors in a machine together by using time multiplexing giving INI slots for every possible connection between any two of the N processors.

The primary computing task for the machine under consideration has a major effect on the network topology. Computing tasks fall into three general categories.
Heavy computational tasks - these tasks require much more computation than network communication. Some examples of this task are pattern recognition (SETI), code breaking, inverse problems, and complex simulations such as weather prediction and hydrodynamics.
Heavy communication tasks - these tasks involve relatively little computation and massive amounts of communication with other processors and with external devices. Message routing is the classic example of these tasks. Other such tasks are data base operations and search.
Intermediate or mixed tasks - these tasks lie between the above or are mixtures of the above. An example of an intermediate task is structured simulation problems, such as battlefield simulation. These simulations require both computation to represent the behavior and properties of the objects (like tanks) and communication to represent interaction between the objects. Some machines may be designed for a mixture of heavy computation and heavy communication tasks.
Historically, supercomputers focused on heavy computation tasks, particularly scientific programming, and mainframes focused on heavy communication tasks, particularly business and database applications.

Read Complete Paper: