IT 专业人士的资源 >
论坛主页
>
Windows HPC Server Job Submission and Scheduling
>
How to use a specific core of a specifc node?
How to use a specific core of a specifc node?
- Hi,
I am new user with Windows HPC cluster. I have a cluster of 7 nodes,each with a quad core processor. I plan to use the cluster for an embarrassingly parallel application i.e. there are several instances of the same application, each of them using different input files and producing different output files. This application is data intensive, having huge input and output files. Hence, I expect there to be a significant bottleneck in the network data access. I plan to design an intermediate application that will automatically send "job" or "mpiexec" commands to the job scheduler.
Is there a way I can submit a task to a particular core of a particular node? That way I can have a finer control, having each core (even of the same node! ) running a different instance of my application, working different input files.
Also, it would be a great plus if you tell me what the advantages are of using the job scheduler rather than using a direct "mpiexec" command, in the case of my scenario.
Thank you very much,
Rohit
答案
- There isn't a way for you to target a specific core, however the scheduler will handle starting up one task (or MPI rank) on each core for you.
If your applicatin is truly "embarssingly parallel" you may not want to use MPI at all. You can simply submit a job with multiple tasks (each of which is an instance of your application).
For example, "job submit /parametric:1-1000 myapp.exe" will run your appliaction 1000 times, with one active process per core in the cluster at any time.
I recommend you check out our docs for more details: http://technet.microsoft.com/en-us/library/cc719020(WS.10).aspx
THanks,
Josh
-Josh- 已建议为答案Josh BarnardMSFT, 所有者：2009年6月9日 17:56
- 已标记为答案Don PatteeMSFT, 版主2009年6月24日 22:11
全部回复
- There isn't a way for you to target a specific core, however the scheduler will handle starting up one task (or MPI rank) on each core for you.
If your applicatin is truly "embarssingly parallel" you may not want to use MPI at all. You can simply submit a job with multiple tasks (each of which is an instance of your application).
For example, "job submit /parametric:1-1000 myapp.exe" will run your appliaction 1000 times, with one active process per core in the cluster at any time.
I recommend you check out our docs for more details: http://technet.microsoft.com/en-us/library/cc719020(WS.10).aspx
THanks,
Josh
-Josh- 已建议为答案Josh BarnardMSFT, 所有者：2009年6月9日 17:56
- 已标记为答案Don PatteeMSFT, 版主2009年6月24日 22:11
