Let $\{\mathsf{T}_t\}_{t>0}$ be a symmetric diffusion semigroup on a $σ$-finite measure space $(Ω, \mathscr{A}, μ)$ and $G^{\mathsf{T}}$ the associated Littlewood-Paley $g$-function operator: $$G^{\mathsf{T}}(f)=\Big(\int_0^\infty \left|t\frac{\partial}{\partial t} \mathsf{T}_t(f)\right|^2\frac{\mathrm{d}t}{t}\Big)^{\frac12}.$$ The classical Littlewood-Paley-Stein inequality asserts that for any $1<p<\infty$ there exist two positive constants $\mathsf{L}^{\mathsf{T}}_{p}$ and $\mathsf{S}^{\mathsf{T}}_{ p}$ such that $$ \big(\mathsf{L}^{\mathsf{T}}_{ p}\big)^{-1}\big\|f-\mathrm{F}(f)\big\|_{p}\le \big\|G^{\mathsf{T}}(f)\big\|_{p} \le \mathsf{S}^{\mathsf{T}}_{p}\big\|f\big\|_{p}\,,\quad \forall f\in L_p(Ω), $$ where $\mathrm{F}$ is the projection from $L_p(Ω)$ onto the fixed point subspace of $\{\mathsf{T}_t\}_{t>0}$ of $L_p(Ω)$. Recently, Xu proved that $ \mathsf{L}^{\mathsf{T}}_{ p}\lesssim p$ as $p\rightarrow\infty$, and raised the problem abut the optimal order of $ \mathsf{L}^{\mathsf{T}}_{ p}$ as $p\rightarrow\infty$. We solve Xu's open problem by showing that this upper estimate of $\mathsf{L}^{\mathsf{T}}_{ p}$ is in fact optimal. Our argument is based on the construction of a special symmetric diffusion semigroup associated to any given martingale such that its square function $G^{\mathsf{T}}(f)$ for any $f\in L_p(Ω)$ is pointwise comparable with the martingale square function of $f$. Our method also extends to the vector-valued and noncommutative setting.