Multicore systems have increasingly gained importance in high performance computers. Compared to the traditional microarchitectures, multicore architectures have a simpler design, higher performance-to-area ratio, and improved power efficiency. Although the multicore architecture has various advantages, traditional parallel programming techniques do not apply to the new architecture efficiently. This dissertation addresses how to determine optimized thread schedules to improve data reuse on shared-memory multicore systems and how to seek a scalable solution to designing parallel software on both shared-memory and distributed-memory multicore systems. We propose an analytical cache model to predict the number of cache misses on the time-sharing L2 cache on a multicore processor. The model provides an insight into the impact of cache sharing and cache contention between threads. Inspired by the model, we build the framework of affinity based thread scheduling to determine optimized thread schedules to improve data reuse on all the levels in a complex memory hierarchy. The affinity based thread scheduling framework includes a model to estimate the cost of a thread schedule, which consists of three submodels: an affinity graph submodel, a memory hierarchy submodel, and a cost submodel. Based on the model, we design a hierarchical graph partitioning algorithm to determine near-optimal solutions. We have also extended the algorithm to support threads with data dependences. The algorithms are implemented and incorporated into a feedback directed optimization prototype system. The prototype system builds upon a binary instrumentation tool and can improve program performance greatly on shared-memory multicore architectures. We also study the dynamic data-availability driven scheduling approach to designing new parallel software on distributed-memory multicore architectures. We have implemented a decentralized dynamic runtime system. The design of the runtime system is focused on the scalability metric. At any time only a small portion of a task graph exists in memory. We propose an algorithm to solve data dependences without process cooperation in a distributed manner. Our experimental results demonstrate the scalability and practicality of the approach for both shared-memory and distributed-memory multicore systems. Finally, we present a scalable nonblocking topology-aware multicast scheme for distributed DAG scheduling applications.