The Logging Option in TSS

Data logging consists of logging the data along with the stripe information in a log device synchronously and then writing to the actual device. After the actual write is complete, a commit of the transaction is done on the log device. The design of the log is depicted in figure 3.

Figure 3: Design of the log

Some of the issues in logging are:

• Should the data logging be only targeting the data consistency or try to improve performance, by delaying the writes to the disks, similar to a log structured file system. The former approach was chosen in TSS, mainly due to its simplicity and ease of addition into the existing TSS device driver.

• At what level should the logging be incorporated in the existing TSS device driver? The possible options were at request level, at stripe level and at physical block level. Considering the advantages and the disadvantages of the three approaches, the stripe level logging was chosen as a suitable method for TSS.

• When should the recovery process be initiated? An ioctl() was provided to trigger the recovery mechanism, which can be called through an application program, in case of crash of the TSS device.

• If the recovery is done in a straight forward manner, using the normal stripe write mechanisms, there is a possibility of data corruption. This case arises due to the use of Read-Modify-Write (RMW) cycles in a RAID5 device. The problem: if a part of a write is committed before the crash, it will leave the stripe in an inconsistent state. If a RMW cycle is used during recovery, then some inconsistent data may be read and used during the recovery that would result in permanent inconsistent data on the disk.

  This problem can be solved, if reconstructing writes are used in place of RMW cycles during the recovery. In reconstructing writes, data is reconstructed from the whole stripe; partially committed data will not affect the recovery and the final data will be consistent.

• What should be done in case of failures in accessing the log device? Logging is used for crash recovery but any redundancy in the log device to avoid the failure of the device is likely to be costly.