search

ut-parla / parla-experimental

5 stars 0 forks source link

`verify_order` is not working correctly #73

Open nicelhc13 opened 1 year ago

nicelhc13 commented 1 year ago

python benchmark/python/benchmark.py -graph sc23_inputs/reduction_6250_2_16000_0_8_3_2.gph -computation_weight 16000 -gil_count 1 -gil_time 0 -user 0 -num_gpus 3 -overlap 1 -level 8 -branch 2 -d 2 -data_move 2 -workloads reduction -width_bytes 6250

wlruys commented 1 year ago

I'll add actual runahead support to this over the weekend. The verifcation codes weren't written in a way thats very compatible with all the different runtime options.

nicelhc13 commented 1 year ago

I saw your PR even though I didn't check all the codes. I thought that this verification code should work for synthetic workloads that runs on the graph written based on our currently using grammar, shouldn't it? Otherwise, what do you mean compatibility? Could you please explain this to me more specifically? If reduction workloads are not compatible, I can remove that from our tests ofc.

wlruys commented 1 year ago

My PR should have fixed the issue you're seeing, but apparently it is still occurring so I'll need to revisit this.

By compatibility I mean if we ever want to run test synthetic apps with runahead (aka PARLA_DEFAULT_SYNC != 0) then the current verification codes need to be rewritten to support those cases.

nicelhc13 commented 1 year ago

I see. Thanks!