Hash keys and the table Pagesize.

Question

Hi,
We look to be running into some performance issues running on TimesTen 5.1.25, and feel that we may be running into a problem of having the hash size set too small, as we are seeing lock contention on the index hash values.
The documentation says that the hash size can only be changed by recreating the table, however the documentation also states, that to resize a hash index, the "ALTER TABLE owner.name SET PAGES = CURRENT;" statement can be used.
Could you please explain which of these statements is correct, and whether this may be leading to greater lock contention, or whether we should be looking elsewhere.
Regards,
mgh

Answer 1

Yes, an undersized hash index can lead to performance issues and excessive contention. One should always size the index based on the formula PAGES = (maximum number of rows) / 256. The ttXactAdmin utility can be used to investigate lock and latch contention on the hash index. The degree of impact depends on the ratio of actual rows to expected rows. If A is the actual number of rows and H is the number of rows for which the hash index has been sized then the ratio A / H should always be <= 1. If it exceeds 1 by a small amount (e.g. 1.1) then the impact is small but once it exceeds one by a significant amount e.g. 1.5 upwards then the impact increases very quickly.
In 5.1, the only way to resize the index is by unloading the table data to a file (ttBulkCp), dropping the table, re-creating with the 'correct' hash index sizing and then reloading the data (ttBulkCp).
The ability to resize the index using ALTER TABLE was introduced in TimesTen 6.0.
Since this is an obvious and well known source of performance problems and increased contention I would recommend that you try to eliminate this issue in the first instance and see if that improves things. If it does not then you can move on to investigate other areas.
Chris