Optimization

To optimize performance and load balancing, the database server data-flow manager partitions the conversion task into separate threads from the read and insert threads and then routes the conversions to coservers that have available resources. Once the rows are converted, the data-flow manager sends each row to the appropriate insert thread.

Figure 77 shows how multiple data-flow managers and conversions tasks balance the load process across multiple coservers:

• A user on coserver_x issues an INSERT statement via DB-Access to load a file from disk and convert its format.
• The data-flow manager discovers that the source file resides on one disk on coserver_x and two disks on coserver_z.
• Readers (sometimes referred to as loader threads) on coserver_x and coserver_z read the file in parallel through named pipes. The database server receives the data through a named pipe.
• The data-flow manager uses a round-robin scheme to distribute the data to all coservers.
• Converter threads (one per CPU VP) on each coserver perform any required data reformatting and pass rows of data back to the data-flow manager.
• The data-flow manager buffers and partitions the data among the coservers that have disks with dbspaces allocated to the table (disk_y5 and disk_z2).
• Insert threads place the data on disk.
  Figure 77. The Load Process