Changes between Version 1 and Version 2 of ticket/0871/Review
- Timestamp:
- 2011-10-19T16:50:50+02:00 (13 years ago)
Legend:
- Unmodified
- Added
- Removed
- Modified
-
ticket/0871/Review
v1 v2 21 21 Ticket, Documentation and Code comments 22 22 23 This development deals deals with tools development, not directly with NEMO code. Policy about tools development is a bit fuzzy, but I guess this code fits NEMO standard with incore documentation, integration inside standard compilation environment.23 This development deals deals with tools development, not directly with NEMO code. Policy about tools development is a bit fuzzy, but I guess this code fits NEMO standard with incore documentation, integration inside standard compilation environment. 24 24 25 25 Short documentation could be added in NEMO documentation (mpp part or specific annex). … … 36 36 || Is this change expected to preserve all diagnostics? || NA || 37 37 || If no, is reason for the change valid/understood? || NA || 38 || Are there significant changes in run time/memory? || YES ||[[BR]] 38 || Are there significant changes in run time/memory? || YES || 39 39 40 40 41 … … 47 48 Various tests have been performed on intel with ifort compiler and IBM with xlf: 48 49 49 - first without openmp : works fine for different kind of files (restart, outputs), if the size of the variable to rebuild doesn't exceed half of the RAM memory. For instance it fails in the case of 50 a 3D variable of dimension 300*300*50 with daily outputs during 1 year, with 8gb of RAM, the size of the vraible is more of 4Gb and has to be allocated twice (the global array and the sum of the local arrays).[[BR]] 51 The old rebuild is allocating such a variable time step per time step and succeed in this case 52 - with OpenMP and ifort : it fails first on execution (declaration of indimlens), after this modification (indimlens declared as SHARED), it runs fine but the rebuild fileds are not correct. Rebuild fields become correct when desactivating the nested OpenMP loops, and using OpenMP only on the number of input files. 53 - with OpenMP and xlf : runs fine for 3D fields, fails on surface fields. I didn't try to find why. In all cases it fails with big variables. 50 - first without openmp : works fine for different kind of files (restart, outputs), if the size of the variable to rebuild doesn't exceed half of the RAM memory. For instance in the case of 51 52 a 3D variable of dimension 300*300*50 with daily outputs during 54 53 55 54 '''Code Review''' … … 64 63 65 64 === Review Summary === 66 67 OpenMP problems doesn't seem to be major drawback. I guess problems with ifort where more related to the support of nesting in the given implementation.[[BR]] 68 I have to re-check for the 2D fields on IBM, if there is a problem, it could be fixed easily since 3D fields are OK. 69 The memory problem is an issue to me, off course depending of the configuration in use. 65 Add summary here 70 66 71 67 === Approval for the trunk ===