Mistakes of Workload calculation

[fly-dragon211]

The calculation of $Mss_{GPU}$ and $Mss_{CPU}$ may have some mistakes. The following is my calculation procedure:

while in the paper:

[shadow150519]

i think they do make the mistake :(

[shadow150519]

The calculation of MssGPU and MssCPU may have some mistakes. The following is my calculation procedure:

while in the paper:

Can you tell me do they build the system using the wrong result or just a slip of the pen ？

[ymjiang]

@shadow150519 Yes, this is a typo in the paper. Despite the typo, the final conclusion in equation (6) is right and BytePS correctly implements the traffic allocation strategy. Thank you for pointing out this.

[shadow150519]

@ymjiang base on tc = tg and eq3, eq4, eq5, i get the following result

$$ M_{ss_{GPU}} = \frac{k - 1}{-kn + 2k -n}*M $$

$$M_{ss_{CPU}} = \frac{2(1-n)}{-kn+2k-n}*M$$

and the t_opt should be the following result

$$ t_{opt} = \frac{2(1-n)}{(-kn+2k-n)*B}*M $$

eq7 is wrong and should be the following

$$ \gamma_{\alpha} = \frac{\frac{2(n-1)M}{nB}}{t_{opt}} = \frac{kn-2k + n}{n} $$

$$ \gamma_{p} = \frac{\frac{nM}{kB}}{t_opt} = \frac{n(-kn+2k-n)}{2k(1-n)}$$

i think these should be the correct result

[shadow150519]

i think you do the calculation totally wrong in section 4.1