Clearly, it is not acceptable for an application to fail due to a crash of another workstation used as a remote memory server. Mirroring is the simplest solution that may be applied but it has a high memory overhead. Parity caching may be used instead, in which case the latency would increase in case of a crash. In both cases the network topology should be taken advantage of in order to reduce the messages needed to achieve reliability. We believe that all three options should be provided (no reliability, mirroring and parity).