Distributed real-time computing system JStrom significantly improves performance on real-time data computing. However, JStorm may lead to a single point of failure, and upgrading JStorm system may cause instability of system. To efficiently solve these problems, we propose a task detection model based on the heartbeat detection and a two-level cluster load balancing model separately. We discover the running state of task through heartbeat detection algorithm, and take action accordingly to avoid single point of failure. The two-level cluster load balancing model divides large cluster into sub cluster, which is isolated and upgraded one by one. Our experimental results suggest that our model can efficiently solve the underlying problems of JStorm system, significantly improve the stability of the system, reduce the probability of single point of failure and the latency of cluster process, and keep system stability during the upgrade process.