Opened 11 years ago

Closed 11 years ago

Last modified 18 months ago

#445 closed Bug (fixed)

Performance of NEMO 3.1

Reported by: ebehrens Owned by: nemo
Priority: high Milestone:
Component: OCE Version: release-3.1
Severity: Keywords: MPI
Cc:

Description

NEMO 3.1 on a NEC SX9 has half of the performance compared to NEMO 2.3. This can be attributed to the communication along the north-fold boundary condition. We have performed the following test:

  1. ORCA05 on 4x1 PE under NEMO 2.3
  2. ORCA05 on 4x1 PE under NEMO 3.1 (with IOF)
  3. ORCA05 on 4x1 PE under NEMO 3.1 jperio=1 (with IOF)

Commit History (0)

(No commits)

Attachments (6)

FTRACE_output.tar (170.0 KB) - added by ebehrens 11 years ago.
Ftrace output for performed Experiments
NEMO_2.3_1215timesteps_FTRACE.txt (60.2 KB) - added by ebehrens 11 years ago.
NEMO_3.1_1215_timesteps_FTRACE_with_IOF.txt (49.0 KB) - added by ebehrens 11 years ago.
NEMO_3.1_jperio1_1215_timesteps_FTRACE_with_IOF.txt (49.8 KB) - added by ebehrens 11 years ago.
NEMO_3.1_1215_timesteps_jperio1_with_IOF_MPI_INFO.txt (4.1 KB) - added by ebehrens 11 years ago.
NEMO_3.1_1215_timesteps_with_IOF_MPI_INFO.txt (4.1 KB) - added by ebehrens 11 years ago.

Download all attachments as: .zip

Change History (13)

Changed 11 years ago by ebehrens

Ftrace output for performed Experiments

Changed 11 years ago by ebehrens

Changed 11 years ago by ebehrens

comment:1 follow-up: Changed 11 years ago by rblod

HI

  • it would be nice to have the global MPI information for both run with NEMO_3.1 too
  • running 3.1 without IOF could be nice also to clarify the situation
  • finally on NEC computers we use to run in cutting only along latitude, both for vectorization and to get rid of the north fold condition

Rachid

Changed 11 years ago by ebehrens

comment:2 in reply to: ↑ 1 ; follow-up: Changed 11 years ago by ebehrens

Replying to rblod:

HI

  • it would be nice to have the global MPI information for both run with NEMO_3.1 too
  • running 3.1 without IOF could be nice also to clarify the situation
  • finally on NEC computers we use to run in cutting only along latitude, both for vectorization and to get rid of the north fold condition

Rachid

Hi

  • MPI INFOS see attachments
  • comparison between IOF and offline are under way

Erik

comment:3 in reply to: ↑ 2 Changed 11 years ago by ebehrens

Hi

-concerning "cutting only along latitude" (1x4) and IOF (extra halos) we have problems, and will submit a new separate ticket after proofing results

Erik

comment:4 Changed 11 years ago by rblod

Hi

I just noticed in a previous mail from Markus than he was using noopt_ieee option for compilation. If it is the case in your benchmark, this option shouldn't be in use and can slow down drastically the code.

Rachid

comment:5 Changed 11 years ago by ebehrens

  • Resolution set to fixed
  • Status changed from new to closed

comment:6 Changed 18 months ago by andmirek

In 10659:

GMED ticket 445

comment:7 Changed 18 months ago by andmirek

In 10660:

GMED ticket 445 remove write statements

Note: See TracTickets for help on using tickets.