The uplink data arriving at the Machine-to-Machine (M2M) Application Server (AS) via M2M Aggregators (MAs) is fairly heterogeneous along several dimensions such as maximum tolerable packet delay, payload size and arrival rate, thus necessitating the design of Quality-of-Service (QoS) aware packet scheduler. In this paper, we classify the M2M uplink data into multiple QoS classes and use sigmoidal function to map the delay requirements of each class onto utility functions. We propose a proportionally fair delay-optimal multiclass packet scheduler at AS that maximizes a system utility metric. We note that the average class delay under any work-conserving scheduling policy can be realized by appropriately time-sharing between all possible preemptive priority policies. Therefore the optimal scheduler is determined using an iterative process to determine the optimal time-sharing between all priority scheduling policies, such that it results in maximum system utility. The proposed scheduler can be implemented online with reduced complexity due to the iterative optimization process. We then extend this work to determine jointly optimal MA-AS channel allocation and packet scheduling scheme at the MAs and AS. We first formulate a joint optimization problem that is solved centrally at the AS and then propose a low complexity distributed optimization problem solved independently at MAs and AS. We show that the distributed optimization solution converges quickly to the centralized optimization result with minimal information exchange overhead between MAs and AS. Using Monte-Carlo simulations, we verify the optimality of the proposed scheduler and show that it outperforms other state-of-the-art packet schedulers such as weighted round robin, max-weight scheduler etc. Another desirable feature of proposed scheduler is low delay jitter for delay-sensitive traffic.