On using network attached disks as shared memory

Abstract

Recent advances in storage technology have enabled systems like Storage Area Networks, where disks are attached directly to the network, rather than being under the control of a single process. In such an environment there is no a priori bound on the number of processes that may access the network attached disks, and so uniform implementations are desirable, that is, implementations that do not rely on the number of processes. We investigate how to use network attached disks, where some disks may crash, as a shared communication medium. To do so, we model disk blocks as Multi-Writer Multi-Reader (MWMR) shared memory registers that may fail by crashing. We study whether a finite number of such fail-prone registers can be used to uniformly implement various types of fail-flee target registers: wait-free atomic, atomic, and wait-free sequentially consistent. For each of these types, we determine the implementability of Multi-Writer Multi-Reader registers, Multi-Writer Single-Reader registers (MWSR), Single-Writer Multi-Reader registers (SWMR) and Single-Writer Single-Reader registers (SWSR). For example, we show that there is no uniform atomic implementation of a MWMR register using finitely many base registers, even if the implementation need not be wait-free. On the positive side we show that with infinitely many base registers then all types of registers can be implemented. This opens the question of how to translate uniform shared memory protocols that use MWMR registers to use network attached disks.