Launching AM Docker ver. 5.7.17 on the new Mac ARM (M1)

[nicolasray]

AccessMod (amd64) eventually runs in Docker (for ARM) installed on the new Macbook Pro ARM with M1 Pro. I understand it's using Rosetta for that.

Here the sequence I got when

1- installing Docker for ARM --> OK
2. launching AccessMod Docker 5.7.17 --> the container installed in Docker, but then the following error in AccessMod App:

3. closing the error windows, then a blank AccessMod windows appears

5. closing AccessMod, and restarting it, now AccessMod opens and runs normally

But is seems quite slower that when running on the ARM64 Macbook Pro, and this is expected. For example, the geographic coverage with parameters by default on the DEMO project takes 1.34 minutes

[fxi]

Hard to reproduce until I get an ARM based mac.

Yep rosetta for the Launcher + qemu for containers. That's quite a drag.

Same analysis on my x86 machine: 0.244 minutes ( 8 CPU, 8 GB RAM, 70GB free disk space)

Potential workaround would be to connect the launcher to a cloud instance, so only electron would have to be translated by roseta2. Users Volume support should still be implemented, though.

[fxi]

5.7.20-alpha should work on ARM

[nicolasray]

I was on 5.7.18, running correctly in Docker, on my ARM M1 macbook Pro. When I launched the update to 5.7.20-alpha from AccessMod, an error message says (something like): "The package was not found"

And then I have the "Loading..." page that stays forever. I closed AccessMod and reopened it, and it's now running version 5.7.20-alpha. I re-ran the geographic coverage with default DEMO parameters. It now runs in 0.37 min, much better than with Roseta interfering and my previous 1.34 minutes (but slower than you 0.244 minutes).

[fxi]

Yep, maybe an issue with docker switching from x86 image to arm image.

For performance, we have some new issues : I've done the geographic coverage again, project demo, 26 seconds.

So, quick recap

[5.7.18]
X86 = 15 s
Server Google cloud, 4GB RAM = 14s

[5.7.20-alpha-1.0]
M1 = 22 s
M1 pro = 26 s
Rosetta + M1 = 1 m 20 s
Server ISE x86 , 72 cpu, 250GB RAM = 37 s
Server Google cloud, 4GB RAM = 35s

-> huge performance regression using the latest image.

Probably caused by:

MYCFLAGS="-O2 -std=gnu99 --march=native" \
    MYLDFLAGS="-s -Wl,--no-undefined -lblas" \

instead of

MYCFLAGS="-O2 -std=gnu99 -m64" \
    MYLDFLAGS="-s -Wl,--no-undefined -lblas" 

but -m64 is not enabled for aarch:

Check also -O3 flag.

https://gcc.gnu.org/onlinedocs/gcc/Option-Summary.html

-> requires some investigation + test if something else ( e.g. grass pseudo session ) could cause this huge gap.

[nicolasray]

@fxi : is this performance regression that was found still an issue with the latest version 5.8.-alpha2 on ARM ? If so this should be put as a Priority 2 and investigated.