IOSS reader is slow with small, parallel dataset
The IOSS reader is being slow reading a dataset with a lot of files, variables, many timesteps, numerous blocks but very, very few cells and points. I am comparing between 5.9.1 and 5.10.0. 5.10.0 is about 5 to 10 times slower. Here is how to replicate:
- I am giving you two files, spreader32 and exampleFile.e. Create a directory work. Move spreader32 and exampleFile.e into work. "source spreader32". You should now see a dataset called exampleDataset.e.32.[00-31].
- 5.10.0, Linux, remote server. I'm using 16 ranks, I don't know what 2 would do.
- 5.9.1, Linux, remote server. I'm using 16 ranks, I don't know what 2 would do.
- With 5.9.1 open exampleDataset.e.32.*. Apply.
- With 5.10.0 open exampleDataset.e.32.*. BE SURE TO TURN ALL VARIABLES OFF. Apply.
- Now, step forward in time with 5.9.1. Count seconds to next timestep. Should be 2 seconds to 5 seconds (on Chama)
- Now, step forward in time with 5.10.0. Count seconds to next timestep. Should be about 30 or more seconds (on Chama)
Spread, we should have around 5,000 cells and points. 100 timesteps. 13 blocks. No variables turned on (but lots of variables, maybe 100 or so).
5.10.0 is about 5 to 10 times slower than 5.9.1. This is a critical bug.
- Just to see how horrible loading all the variables is, on 5.10.0, turn all the variables on and hit apply. On Chama this takes minutes.
I will send exampleFile.e and spreader32 to Utkarsh.
OK to slip to 5.11 if the fix isn't easy or safe.