Parallel Fluent Processing Wang Junhong SVU/Academic Computing, Computer Centre 1. What is parallel processing of Fluent? − Parallel processing of Fluent means to run Fluent solver on two or more CPUs simultaneously to calculate a computational fluid dynamics (CFD) job. 2. Why need to use parallel processing of Fluent? − − What are the computing resources and how about the parallel performance? − Todate, there have 3 different parallel computing resources and 5 batch queues in SVU for users to run parallel processing of Fluent: Compute Server Compaq (cheetah, cheetah2) Linux Cluster (atlas) Itanium2 Linux (cougar1~5) − Batch Queue Max. # of CPU MEM Limit cpq_3p 3 2 GB cpq_8p cpq_8p_8gb 8 4 GB 8 GB linux_4p 4 4 GB ia64_4p 4 4 GB Benchmark performance for parallel processing of Fluent on the above servers and batch queues is illustrated as below: 1000 900 Computing Time (min) 3. Reduce computing time: splitting a job to two or more small partitions, hence to take less time to complete and thus cut down the time to solution. Make large scale job doable: large scale job that has no way or, if not, impossible to be processed on a single CPU due to the restriction of hardware (i.e., RAM , disk space) and time, can become doable after being segmented to many small partitions which could then be handled by many CPUs. 857.0 800 700 600 500 364.8 400 300 225.3 200 100 0 98.8 47.7 1p 3p 112.3 34.9 8p 1p Compaq (ch, ch2) 4p Linux Cluster (atlas) 1 1p 4p Itanium2 (cougar1-5) o o o o 4. The computing time in the above chart is the elapse time (or wall clock time) needed for the job to complete in the individual queue. The benchmark performance very clearly indicates that you can cut down the computing time tremendously if run Fluent in parallel. However, the real speedup rate of your simulation may be different from what is listed here. It will be much depended on the system load, status and other factors during the specific period your jobs are being processed. The test case used for the benchmarking is a mixing problem, with k-epsilon turbulence model, heat transfer and about 800,000 mesh cells in the computational domain, takes about 1 GB of memory. How and where to submit/run parallel Fluent jobs? − − − Fluent is a well parallised CFD solver. Users don’t have to know parallel programming knowledge and don’t need to write parallel codes in order to run parallel Fluent simulation. Fluent will automatically partition the job for users, using defaults that should be close to optimal. Parallel Fluent jobs need to be submitted to batch compute queues and run in batch mode in SVU. All jobs are managed by LSF manager The submission command to 5 parallel queues are: Queue Sample Command (for 3d case) cpq_8p bsub –q cpq_8p –o job1.out “mpiclean; fluent 3d –t8 -pvmpi –g –i job1.script; mpiclean” linux_4p bsub –q linux_4p –o job2.out –n 4 fluent 3d –t0 –pnet –g –i job2.script –lsf ia64_4p bsub –q ia64_4p –o job3.out “fluent 3d –t4 –psmpi –g –i job4.script” * note: the command for cpq_3p and cpq_8p[_8gb] is similar. − − − 5. The 3 sample commands will submit a parallel Fluent job to run on 8, 4 and 4 CPUs at queues cpq_8p, linux_4p and ia64_4p, respectively. You can submit batch Fluent job from any host in SVU, except for linux_4p that requests to logon to atlas. Please visit SVU page http://www.nus.edu.sg/comcen/svu/techinfo/fluent_cpqpll.html and http://www.nus.edu.sg/comcen/svu/techinfo/faq_cfd.html for more information. Tips − − − − − Check queue status (i.e., how many pending jobs) before submit job to the queue. To minimize the waiting time, don’t stick to one queue if many jobs are pending in the queue, even it is faster than others. Parallel processing of Fluent is not recommended to use for the initial test of problem setup or small jobs that only runs a few minutes. Enable checkpoint or set auto backup of the intermediate solution to minimize the impact in case the server is down due to some unforeseen reason. Watch out SVU notice closely of any changes/updates of the systems. Don’t hesitate to email [email protected] to get help and support from us. 2
© Copyright 2024 Paperzz