Checkpoints in Gem5-GPU and counting ticks

[Daniel Gerzhoy]

Hi all,

Say I want to create a checkpoint

The help options say:

--take-checkpoints=TAKE_CHECKPOINTS <M,N> take checkpoints at tick M and every N ticks thereafter

so for example would it be:

...gem5.opt whatever.py --take-checkpoints=<3000000,1000000> ...etc

Start at 3million ticks and take one every million ticks. Right? Or is it a different syntax

Also, how do I count the ticks of gem5 vs the ticks in gpgpusim? are they exactly the same? 

Where is the particular tick variable that the checkpoint option interacts with?

Thanks,

Dan

[Jason Lowe-Power]

Hi Dan,

First, let me say what I usually do for checkpoints. All of my applications are annotated with region of interest (ROI) begin/end magic instructions. I then run gem5 with the --work-begin-exit-count=1 option, which exits simulation at the beginning of the ROI. I also specify --checkpoint-at-end which checkpoints when the simulation exits. With these options, I can generate a checkpoint at the beginning of the ROI.

Second, as a general statement, checkpointing is controlled by the python config scripts. You just need to call m5.checkpoint(<path>) any time simulation is not running (e.g., after m5.simulate()).

Finally, to answer your specific question, I think the syntax for the take-checkpoints option is just --take-checkpoints=1000,10. See (around) line 230 in gem5/configs/common/Simulation.py. 

GPGPU-Sim does use the same ticks as gem5. It is a clocked object just like everything in gem5. See the GPGPUSimComponentWrapper object in gem5-gpu/src/gpu/gpgpu-sim/ for how exactly this works.

Just FYI, gem5-gpu does not support checkpointing during kernel execution (in fact, it may or may not work after any kernel has been run). This is why I only use checkpointing at the start of the ROI, before any kernels have executed. I doubt you want to use periodic checkpointing (--take-checkpoints). I would either use an ROI, or specify an end tick (-m) and checkpoint at the end.

Let us know if you have more questions.

Cheers,

Jason