Question: about ngs/partSize choosing

[initzhang]

Hi Yuke, thanks for your excellent work!

I have a question about the selection of ngs parameter. In the Section 6 of the paper, I only find some constrains for the ngswithout direct formula. In the code repo, I find that in auto model ngs = avgNodeDegree, is this the final criteria of ngs?

https://github.com/YukeWang96/OSDI21_AE/blob/f129823ae49f3b557ef525aaa189fc5c703e5c59/GNNAdvisor/param.py#L73

In the settings I of Fig 14, the black dot gives ngs=1024 for amazon0505, but the average degree is about 12 for this dataset. I am confused about this, could you please tell me where the problem is? Thank you!

[YukeWang96]

Hi, Zhang,

Thanks for your interest of our work and reaching out to us about your questions!

• The runtime configuration setting can be determined manually or automatically determined. https://github.com/YukeWang96/OSDI21_AE/blob/f129823ae49f3b557ef525aaa189fc5c703e5c59/GNNAdvisor/GNNA_main.py#L34
• Thanks for pointing this out to us, and we will make the corresponding correction later on. For searching the parameters in Fig.14, I recommend to follow the formula about WPT in Equ.5 and the discussion of searching process at left column of page 10, which would be a better way to determine runtime parameters (e.g., partSize).

[initzhang]

Hi Yuke, thanks for the reply!

(1) Are the auto-selected results in the paper obtained by setting manual_mode==False in the code? Or we actually need to calculated the best hyper parameter first, and then assign the values through arguments manually?

(2) Could you please elaborate on why there is a constraint of WPT ≈ 1024? Is this a heuristic estimation based on the experiments/benchmarking?

[YukeWang96]

Hi, Zhang, Thanks for following up!

• For this public code release, we mainly demonstrate the kernel design and kernel-level runtime design configurability (decider part is outdated, sorry). Therefore, it is necessary to follow the steps described at Section 6 for the search illustrated in Figure 14.
• Yes, it is a heuristic based on our experiment study based on a given hardware (e.g., GPU/CPU) setup.

[initzhang]

Got it, Thank you!