Soft decision decoding of RAID stripe for higher endurance of flash memory based solid state drives

Abstract

RAID schemes have been in use for hard disk drives to provide improved performance or fault tolerance. RAID schemes, particularly RAID 2 and higher comprise of striping and adding parity. For solid state drives (SSDs), to combat die failures, schemes like RAID 5 and RAID 6 are effective. The idea is to have block level striping with distributed parity to provide data recovery in case of single or multiple die failures. In this work, we analyze the RAID 5 case used in SSDs when there is no die failure. For this case, the distributed parity information can be used to either combat higher raw bit error rate (RBER) or to improve the uncorrectable bit error rate (UBER). The algorithm we propose can be used with BCH or LDPC codes. With LDPC codes, it uses soft information on RAID 5 parity in order to recover from multiple error correcting code (ECC) fatals in a RAID stripe. The entire stripe data is read and then decoding is performed using the RAID parity to try recovery. We used simulations in order to characterize the RBER gains. First, modeling of the RBER distribution of the dies was performed using the log-normal distribution. Using this model, a performance evaluation for ECC schemes like LDPC codes and BCH codes was performed. We show that if LDPC codes are used, RAID 5 theoretically offers an RBER capability which is 3 times that of the error correcting code RBER capability. However, it is not possible to realize this RBER benefit completely since it has SSD performance downside. In order to meet the Quality of Service (QoS) specification, a realizable RBER gain is by a factor of 1.3 to 2 instead of 3. Further, these gains can only be realized if an LDPC code is used as an error correcting code, algebraic codes like BCH codes will provide limited gains. If no RBER benefit is permitted by a particular configuration of solid state drive, this scheme can be used to provide UBER gain by 3 orders of magnitude.