By integration of
$$\frac1{\sqrt x}\le\frac1{\sqrt{\lfloor x\rfloor}}<\frac1{\sqrt{x-1}}$$
we establish a formula for the tail of the summation,
$$\int_m^{n+1}\frac{dx}{\sqrt x}\le\sum_{k=m}^n\frac1{\sqrt k}<\int_m^{n+1}\frac{dx}{\sqrt{x-1}},$$ or
$$2(\sqrt{n+1}-\sqrt m)\le\sum_{k=m}^n\frac1{\sqrt k}<2(\sqrt n-\sqrt{m-1}).$$
Then for the complete sum
$$\sum_{k=1}^{m-1}\frac1{\sqrt k}+2(\sqrt{n+1}-\sqrt m)\le\sum_{k=1}^n\frac1{\sqrt k}<\sum_{k=1}^{m-1}\frac1{\sqrt k}+2(\sqrt n-\sqrt{m-1}).$$
We can compute the bracketing for increasing $m$, and it turns out that $m=2$ suffices to establish
$$1998.17257288\le S<1999.0$$
Hence, $$\color{green}{1998}.$$
For safety, with $m=3$,
$$1998.24400517\le S<1998.87867966$$
Rationale of the method:
An integral approximates a sum inasmuch as the function value remains sufficiently constant in unit intervals. In the case of the inverse square root, the error per interval is
$$\frac1{\sqrt{n+1}}-\frac1{\sqrt n}=O(n^{-3/2}),$$ which decreases relatively quickly. So there is some hope that by combining a direct summation of the first terms and a bracketing of the tail, we can end up with the two bounds in the same unit interval so that we know the integer part. (The last terms won't be a problem as they correspond to a tiny error.)
We are pretty lucky here as the unit interval is found with very few terms.
Given that the width of the bracketing is
$$2(\sqrt n-\sqrt{n+1}+\sqrt m-\sqrt{m-1}),$$ we can conclude as soon as the partial sum differs from an integer by more than this amount. And the firt parial sums are approximately
$$1,1.71,2.28,2.78,3.23\cdots$$
which leave a good margin.