To meet the ever-growing mobile data traffic, network spatial densification
with various low-power nodes in addition to the conventional high-power macro base
stations (BS), a.k.a. heterogeneous network (HetNet), is regarded as one key
enabling solution. Due to the unplanned nature, HetNets are very irregular and
severe interference can happen without judicious designs of the user association
rules. Conventional maximum downlink (DL) signal-to-interference-plus-noise ratio
(SINR) association rule, on the other hand, can not fully release the traffic
offloading capabilities of HetNets. Most prior work assumed best-effort (BE)
traffic and sought the optimal association to maximize metrics such as the sum of
log-rate. In this paper, we model the DL quality-of-service (QoS) flow to each
mobile station (MS) explicitly. Aiming to minimize the network-wide packet delay,
we investigate the optimal user association scheme and the corresponding
resource allocation algorithm. Our proposed traffic-aware user
association strategy enjoys low complexity and fast convergence, which is
also corroborated with simulations.