Online auction markets play increasingly important roles for resource allocations in distributed systems. This paper builds upon a market-based framework presented by Guo et al. (Guo, Z., G. J. Koehler, A. B. Whinston. 2007. A market-based optimization algorithm for distributed systems. Management Sci. 53(8) 1345-1458), where a distributed system optimization problem is solved by self-interested agents iteratively trading bundled resources in a double auction market run by a dealer. We extend this approach to a dynamic, asynchronous Internet market environment and investigate how various market design factors including dealer inventory policies, market communication patterns, and agent learning strategies affect the computational market efficiency, market liquidity, and implementation. We prove finite convergence to an optimal solution under these various schemes, where individual rational and budget-balanced trading leads to an efficient auction outcome. Empirical investigations further show that the algorithmic implementation is robust to a number of dealer and agent manipulations and scalable to larger sizes and more complicated bundle trading markets. Interestingly, we find that, though both asynchronous communication and asymmetric market information negatively affect the speed of market convergence and lead to more agent welfare loss, agents' ability to predict market prices has a positive effect on both. Contrary to conventional wisdom that a dealer's intertemporal liquidity provisions improve market performance, we find that the dealer's active market intervention may not be desirable in a simple market trading environment where an inherent market liquidity effect dominates, especially when the dealer owns a significant amount of resources. Different from the traditional market insight, our trading data suggest that high trading volume does not correlate to low price volatility and quicker price discovery.