Reliable Storage Systems through LDPC Array Codes

Speaker: Dr. Yuval Cassuto , ALGO - EPFL


Two-dimensional Array Codes are a central tool to efficiently maintain data reliability and availability in commercial storage systems. Aggregating multiple storage devices into a large pool of storage requires protection from failures of individual devices. The effectiveness of Array Codes in that task is thanks to their two main features:

1) Column based error/erasure model → resilience to massive failures in individual devices

2) Low-density encoding matrices → low complexity encoding, decoding and updates

When the number of device failures grows, the complexity of known algebraic Array-Code families quickly becomes prohibitive for practical implementation.

In the talk we will discuss different frameworks and constructions to overcome this complexity blow-up while increasing the failure resilience of the codes. An important ingredient of the work are new error models that thrust the correction capability toward more realistic failure events from the world of practice.

In the main part of the talk, constructing Array Codes as Low-Density Parity Check (LDPC) codes over a special channel is shown to offer increased flexibility in trading-off correctability with complexity.

The talk will combine practical insight with theoretical reasoning, attempting to address interests of a broad audience.