As one of NOSQL data stores, a document-oriented data store manages data as documents in a scheme-less manner. Various string match queries, such as a perfect match, begins-with (prefix) match, partial match, and regular expression based match, are performed for the documents. To accelerate such string match queries, we propose DistGPU Cache (Distributed In-GPU Data Cache), in which data store server and GPU devices are connected via a PCI-Express (PCIe) over 10 Gbit Ethernet (10 GbE), so that GPU devices that store and search documents can be added and removed dynamically. We also propose a partitioning method that distributes ranges of cached documents to GPU devices based on a hash function. The distributed cache over GPU devices can be dynamically divided and merged when the GPU devices are added and removed, respectively. We evaluate the proposed DistGPU Cache in terms of regular expression match query throughput with up to three NVIDIA GeForce GTX 980 devices connected to a host via PCIe over 10 GbE. We demonstrate that the communication overhead of remote GPU devices is small and can be compensated by a great flexibility to add more GPU devices via a network. We also show that DistGPU Cache with the remote GPU devices significantly outperforms the original data store.