We introduce a variant of the capacitated vehicle routing problem that is encountered in sensor networks for scientific data collection. Consider an undirected graph $G=(V \cup \{\mathbf{sink}\},E)$. Each vertex $v \in V$ holds a constant-sized reading normalized to 1 byte that needs to be communicated to the $\mathbf{sink}$. The communication protocol is defined such that readings travel in packets. The packets have a capacity of $k$ bytes. We define a {\em packet hop} to be the communication of a packet from a vertex to its neighbor. Each packet hop drains one unit of energy and therefore, we need to communicate the readings to the $\mathbf{sink}$ with the fewest number of hops. We show this problem to be NP-hard and counter it with a simple distributed $(2-\frac{3}{2k})$-approximation algorithm called {\tt SPT} that uses the shortest path tree rooted at the $\mathbf{sink}$. We also show that {\tt SPT} is absolutely optimal when $G$ is a tree and asymptotically optimal when $G$ is a grid. Furthermore, {\tt SPT} has two nice properties. Firstly, the readings always travel along a shortest path toward the $\mathbf{sink}$, which makes it an appealing solution to the convergecast problem as it fits the natural intuition. Secondly, each node employs a very elementary packing strategy. Given all the readings that enter into the node, it sends out as many fully packed packets as possible followed by at most 1 partial packet. We show that any solution that has either one of the two properties cannot be a $(2-ε)$-approximation, for any fixed $ε> 0$. This makes \spt optimal for the class of algorithms that obey either one of those properties.