Modern RDMA-enabled data centers face significant performance challenges when handling mixed traffic workloads containing both latency-critical short flows ( \(<100\,KB\) ) and bandwidth-intensive long flows ( \(\ge 100\,KB\) ), as existing load balancing mechanisms predominantly optimize for large flows while short flows suffer from suboptimal performance due to limited rerouting opportunities and single-path transmission constraints. This paper presents Short-flow Multi-path Adaptive Routing (SMAR), a novel flow-size-aware load balancing framework that introduces dedicated optimization mechanisms for short flows through dynamic traffic classification using a dynamic threshold and specialized multi-path transmission strategies. Using a congestion factor based on queue length statistics to determine the selection of rerouting target paths, the system employs round-robin load distribution between two paths with similar congestion characteristics to minimize queuing delays and enhance resource utilization for short flows. The framework operates through sender-side enhancements while maintaining compatibility with existing packet reordering mechanisms by adhering to the two-path constraint that ensures manageable packet reordering complexity, and performance evaluation demonstrates substantial improvements in short flow completion times and overall network utilization while preserving system stability under dynamic traffic conditions.