Multimem: Retrofitting System Availability via a Lightweight Binary Translation Framework

As the size of memory in servers becomes larger and larger, the availability of them is under big pressure as memory failures are common. Memory failures may cause data lose and programs even systems corrupted. This could be a disaster in systems which are occupied in military and finance area. To improve the memory availability, some solutions try to mitigate the occurrence of memory errors while most strategies focus on memory failure tolerance. Hardware solutions like mirror memory needs expensive peripheral equipments while existing software approaches are somewhat complicated and limited by the high overhead in practical usage. In this paper, we present a novel lightweight binary translation framework called Multimem to improve system memory access availability. It is a software approach achieving hardware mirror memory feature via static binary translation technology. Multimem switches native systems to high available systems with two or more copies of memory so when memory failures happen, systems could recover the data from the replica. It is flexible to configure systems availability level as the higher level will be, the larger number of memory backups will be offered. The instruction level memory synchronization reduces backup overhead and lowers the probability of data loss compared with traditional software approaches. The results show that Multimem brings 33% overhead at most.