Inhibitors to High-Performance Sort on the Mainframe

Mainframes have several limitations that keep them from matching the results described earlier. Before delving deeper, here is a review of the key elements that made this record-breaking sort possible:

· Multiple processors

· Multiple disks

· High bandwidth between processors and disks

· Large amounts of memory

· Software that can take advantage of all the above

For all of the hardware elements listed above, hardware for the mainframe costs several times more than equivalent hardware for the Windows Server operating system. The difference in cost is partly historical, but it also is due to the more complex infrastructure that has developed for mainframes.

Due to the high costs associated with mainframe environments, companies do not necessarily have the option of adding more hardware to solve the problem of finding batch windows long enough to run large sorts. Instead, companies must try to manage system resources to optimize sorts without compromising the performance of other applications. This practice leads to a series of tradeoffs in processor load, paging activity, I/O activity, disk utilization, and elapsed time.

Ways to improve mainframe sort performance include DFSORT features such as:

· Hipersorting, which uses Hiperspaces for sorting. Hiperspaces use expanded or central storage and can significantly speed up a sort process. However, Hipersorting uses a lot of CPU bandwidth, and the increased paging results in a total reduction of system performance.

· Dataspace sorting allows sorting to occur completely within main storage, and thus eliminates the need to write intermediate data to disk. The limitation of this method is that a dataspace size limit is usually about 2 GB, which precludes the sorting of large files. This method also incurs a large paging expense.

· Cache Fast Write allows data to be read and written to and from work data sets at cache speed.

· Sequential disk striping can also improve elapsed time performance.

Of course, building the hardware infrastructure is only part of the task; the other part is building the sorting application. DFSORT cannot take advantage of parallel processors, which limits its ability to scale to the levels of a Windows Server–based solution. Hardware that supports Windows Server and the Windows Server software architecture have evolved together so that sorting across multiple processors can be readily achieved.

In summary, achieving record-breaking performance from mainframes is unlikely because of:

· Expense

· The effort involved

· Systems architecture limitations

· Sort utilities limited to single processors