Commercial cloud providers, e.g., Amazon EC2, offer the volume discount for large instance reservation in a time slot, and the majority of cloud jobs are delay-tolerant and do not need to be processed intermittently. These two features create an opportunity for the cloud brokerage service which aggregates and schedules cloud users' rental requests to earn volume discounts from cloud providers and sell to cloud users at a cheap price. A challenge for the broker is to properly schedule delay-tolerant jobs in order to maximize the volume discount amount over time. The scheduling idea is to generate several job bundles so each job bundle can get discount. In this paper, we discuss this problem from the homogeneous model first, where each job has the same processing time and delay-tolerant time, and we propose a dynamic programming approach. Then, we extend the model into the heterogeneous model, where the job processing time and the job deadline can be arbitrary values. In the heterogeneous scenario, we prove that the proposed problem is NP-hard even when the job processing time is unit. Then, we propose a greedy approach which turns out to have an approximation of O(\ln n), where n is the total job number. Extensive trace-driven experiments from Google cluster trace demonstrates that our schemes achieve good performances.