FAQ - CLUSTERTECH Parallel Environment (CPE)

• If your Monte Carlo (MC) simulation takes too long to run, use the CPE MC library to parallelize it.
• If your Finite Difference (FD) computation takes too long to run, use the CPE FD library to parallelize it.
• If you want to port a sub-routine that runs too long to a parallel version, and have a middle-ware that launches and control the parallelexecution conveniently.
  
  

For Monte Carlo simulation, you get linear scaling, near perfect parallelism.

For Finite Difference computation, the scaling depends on the latency and speed of the network, the size of the grid to use, and the amount of computation in each time step. Typically, you get linear scaling from using 2 CPU-core up to a certain limit after which the scaling becomes sub-linear and then level off afterward. If your network has low latency and high speed (e.g. Infiniband, Myrinet), the size of the grid is large, and the amount of computation in each time step is large, the linear scaling region goes up for a large number of nodes / CPU-core.

Yes. CPE ParaConnect tranfers data between the controlling program on desktop computer and the parallel program on the cluster nodes on-the-fly. It does not write data to disk file nor transfer files between disks of different machines.

Yes. CPE provides API, libraries and facilities for you to port your serial program to a parallel program. Specifically, you only port the most computationally-intensive part of your program to a parallel version, and then use the ParaConnect API to launch the parallel program and retrieve results.

If your application is Monte Carlo Simulation or Finite Difference computation, use CPE's MC and FD API to port the program to a parallel version easily. Read our tutorial in the developer's guide for code porting. It is not difficult.

Hotline

• China (Tollfree) 400 8108832
• Hong Kong (852) 2655 6108