|
|
|
Similar to the strategy used in transaction log flushing, the FairCom Server can avoid excessive flushing of data and index files under transaction control. Two additional keywords affect this behavior:
COMPATIBILITY TDATA_WRITETHRU and COMPATIBILITY TINDEX_WRITETHRU force transaction controlled data files and index files, respectively, to be written directly to disk (whenever c-tree determines that they must be flushed from the c-tree buffers), and the calls to flush their OS buffers are skipped. These keywords cause transaction controlled files to be written through the OS file system cache, rather than written into the file system cache and later explicitly flushed at a database checkpoint. This behavior allows I/O costs to be evenly amortized, reducing the amount of I/O that must be done at a database checkpoint. This results in a smaller variance in transaction times, and potentially greater total transaction throughput. For scenarios without heavy and continuous write activity, this alternate behavior frequently results in reduced throughput.
TDATA_WRITETHRU uses the file system cache. The file is placed into a mode that causes the write to go to file system cache and then to disk before returning; the data still resides in file system cache. Compare to UNBUFFERED_IO, which completely avoids the file system cache.
See Also
