Efficient many-to-one communication for a distributed RAID

Abstract

Any set of autonomous workstations, however networked (by a LAN, a MAN, or wireless), can be seen as a collection of networked low cost disks. Such a collection can be operated by proper software so as to provide the abstraction of a single, larger block device, made available to all the participants on a peer-to-peer basis. By adding enough data redundancy, the disk collection as a whole could act as single distributed RAID, providing capacity and reliability along with the convenient price/performance typical of commodity hard disks. This paper reports about issues of communication performance in a prototype of distributed RAID device called DRAID. DRAID offers storage services under a single I/O space (SIOS) block device abstraction. The SIOS feature implies that the storage space is accessible through each of the participant stations, rather than through one or few fixed end-points. The paper focuses on the inefficiency of communication when a client reads data stripes from a number of remote servers in a gigabit Ethernet LAN. The congestion caused by such many-to-one communication pattern has been faced in multiple ways, but the best result has been obtained by modifying the traditional, and unsuccessful, congestion avoidance policy of TCP/IP.