Several Java-based distributed in-memory systems have been proposed in the literature, but most are not aiming at graph applications having highly concurrent and irregular access patterns to many small data objects. DXRAM is addressing these challenges and relies on DXMem for memory management and concurrency control on each server. DXMem is published as an open-source library, which can be used by any other system, too. In this paper, we briefly describe our previously published but relevant design aspects of the memory management. However, the main contributions of this paper are the new extensions, optimizations, and evaluations. These contributions include an improved address translation which is now faster compared to the old solution with a translation cache. The coarse-grained concurrency control of our first approach has been replaced by a very efficient per-object read-write lock which allows a much better throughput, especially under high concurrency. Finally, we compared DXRAM for the first time to Hazelcast and Infinispan, two state-of-the-art Java-based distributed cache systems using real-world application-workloads and the Yahoo! Cloud Serving Benchmark in a distributed environment. The results of the experiments show that DXRAM outperforms both systems while having a much lower metadata overhead for many small data objects.