Difference between revisions of "Improving MPI communication latency on euroben kernels"

From SNIC Documentation
Jump to: navigation, search
 
(6 intermediate revisions by 3 users not shown)
Line 1: Line 1:
 
{{project info
 
{{project info
 
|description=Improving the MPI collective performance by network aware communication  
 
|description=Improving the MPI collective performance by network aware communication  
 +
|fields=Computational science
 
|financing=PRACE
 
|financing=PRACE
|active=Yes
+
|active=No
 
|start date=2010-07-01
 
|start date=2010-07-01
 
|end date=2012-06-30
 
|end date=2012-06-30
Line 10: Line 11:
 
The goal of this project is to improve the performance of some euroben kernels. This project is part of PRACE-1ip, WP7.5. The euroben kernel is a set of synthetic benchmark codes which is developed by PRACE. This kernel is used for benchmarking PRACE systems. Some of these benchmarks, e.g., mod2f uses MPI_Alltoall based collective for doing fast Fourier transform. Generally the MPI_Alltoall scaling on many nodes is not good. The performance of alltoall operation can be improved by introducing network awareness in alltoall operation. This can be done by writing a modified alltoall and then calling that from the application code. We tested our modified alltoall on mod2f benchmark with good results
 
The goal of this project is to improve the performance of some euroben kernels. This project is part of PRACE-1ip, WP7.5. The euroben kernel is a set of synthetic benchmark codes which is developed by PRACE. This kernel is used for benchmarking PRACE systems. Some of these benchmarks, e.g., mod2f uses MPI_Alltoall based collective for doing fast Fourier transform. Generally the MPI_Alltoall scaling on many nodes is not good. The performance of alltoall operation can be improved by introducing network awareness in alltoall operation. This can be done by writing a modified alltoall and then calling that from the application code. We tested our modified alltoall on mod2f benchmark with good results
  
status
+
Status
  
 
* Started code modification. The basic algorithm will remain same. Will map the communication such that there is more communication within node than outside. [30/03/2011]
 
* Started code modification. The basic algorithm will remain same. Will map the communication such that there is more communication within node than outside. [30/03/2011]
Line 17: Line 18:
  
 
* The tests are run on Curie. The standard MPI_Alltoallv calls are repalced by topology aware MPI_Alltoallv_tuned calls. [16/01/2012]
 
* The tests are run on Curie. The standard MPI_Alltoallv calls are repalced by topology aware MPI_Alltoallv_tuned calls. [16/01/2012]
* Writing final report. [28/02/2012]
+
 
 +
* Contributed to deliverable D7.5: [[media:7.5.D_Guidelines_for_D7.5.pdf|report]] [28/02/2012]  
 +
* Writing final report
  
 
== Members ==
 
== Members ==

Latest revision as of 10:52, 23 April 2013

Name Improving MPI communication latency on euroben kernels
Description Improving the MPI collective performance by network aware communication
Project financing   PRACE
Is active No
Start date 2010-07-01
End date 2012-06-30

The goal of this project is to improve the performance of some euroben kernels. This project is part of PRACE-1ip, WP7.5. The euroben kernel is a set of synthetic benchmark codes which is developed by PRACE. This kernel is used for benchmarking PRACE systems. Some of these benchmarks, e.g., mod2f uses MPI_Alltoall based collective for doing fast Fourier transform. Generally the MPI_Alltoall scaling on many nodes is not good. The performance of alltoall operation can be improved by introducing network awareness in alltoall operation. This can be done by writing a modified alltoall and then calling that from the application code. We tested our modified alltoall on mod2f benchmark with good results

Status

  • Started code modification. The basic algorithm will remain same. Will map the communication such that there is more communication within node than outside. [30/03/2011]
  • Some initial results are obtained and some improvements are achieved [25/05/2011]
  • The tests are run on Curie. The standard MPI_Alltoallv calls are repalced by topology aware MPI_Alltoallv_tuned calls. [16/01/2012]
  • Contributed to deliverable D7.5: report [28/02/2012]
  • Writing final report

Members

 CentreRoleField
Chandan Basu (NSC)NSCApplication expertComputational science