The DBS Control utility displays and allows modification of various Teradata Database database-level configuration parameters. Several of these can be used to tune memory management and in that way affect system performance.
DictionaryCacheSize Performance Implications
The default value allows more caching of table header information and reduces the number of I/Os required. It is especially effective for workloads that access many tables (more than 200) and for those that generate many dictionary seeks.
Increase the size of the dictionary cache to allow the parser to cache additional data dictionary and table header information.
For tactical and Online Complex Processing (OLCP) type workloads, maintaining a consistently short, few-second response time is important. These workloads may benefit from a larger dictionary cache, particularly when their query plans have not been cached in the request cache. A larger dictionary cache will allow more dictionary detail, needed for parsing and optimizing, to remain in memory for a longer period of time. For query workloads with a response time of more than one minute, there may be no measurable difference when this field is set to a higher value.
IdCol Batch Size Performance Implications
The IdCol Batch Size setting involves a trade-off between insert performance and potential gaps in the numbering of rows inserted into tables that have identity columns.
A larger setting results in fewer updates to DBC.IdCol in reserving batches of numbers for a load. This can improve the performance of bulk inserts into an identity column table. However, because the reserved numbers are kept in memory, unused numbers will be lost if a database restart occurs, resulting in a gap in the numbering of identity columns.
PPICacheThrP Performance Implications
Under most use cases, the default value for PPICacheThrP is adequate, and should not be changed. However, if there are performance issues that might be addressed by adjusting this value, consider the information in this section.
The current data block for the corresponding partition (or buffer, in the case of inserts to column-partitioned tables) is associated with each context. The current set of data blocks or buffers (one for each context) are kept in memory, if possible, to improve the performance of processing the set of partitions at the same time. If there is a shortage of memory, these blocks or buffers may need to be swapped to disk. Excessive swapping, however, can degrade system performance.
Larger values may improve the performance of partitioning operations, as long as the following occur:
- Data blocks or AMP buffers for each context can be kept in memory. When they can no longer be kept in memory and must be swapped to disk, performance may degrade.
- The number of contexts does not exceed the number of nonempty, noneliminated partitions for partitioning operations. (If they do, performance will not improve because each partition can have a context, and additional contexts would be unused.)
In some cases, increasing the value of PPICacheThrP above the default value can provide a performance improvement for individual queries that do these partitioning operations. However, be aware of the potential for memory contention and running out of memory if too many of these queries are running at the same time.
The default setting of 10 is conservative, and intended to avoid such memory problems. With 80 AMP Worker Tasks (AWTs) per AMP on a system with the default setting of 10, the maximum amount of FSG cache that could be used for these partitioning operations is 80% of FSG cache memory, if all AMPs are simultaneously executing partitioning operations, such as sliding-window joins for 80 separate requests. For configurations that have more than 80 AWTs defined as the maximum, the setting is scaled to the number AWTs. For example, at the default setting of 10, a cap of 80% of FSG cache memory per AMP would still be in effect on such systems.
For many sites, the default may be too conservative. All 80 AWTs might not be running partitioning operations at the same time. If, at most, 60 partitioning operations are expected to occur at the same time, the value of PPICacheThrP could possibly be raised to 15. If at most 40 are expected, the value could possibly be raised to 20, and so on. The best value for this parameter is dependent on the maximum concurrent users expected to be on the system and their workload. No one value is appropriate for all systems.
Also, consider that the number of concurrent partitioning operations, such as sliding-window joins, may increase as partition usage is increased. Increasing the value may increase performance now without memory contention or running out of memory but, in the future, as more partitioning operations run concurrently, performance may decrease, or out of memory situations may occur.
If less than 80 concurrent partitioning operations are expected for your site, and you think that better performance may be possible with an increased value for PPICacheThrP, you can experiment with PPICacheThrP settings to determine an increased PPICacheThrP setting that is optimal for your site and safe for your workloads. Measure pre- and post-change performance and degree of memory contention under expected current and future workloads to evaluate the effects of the change. If increasing the value to one that is reasonably safe for your site does not yield adequate performance for partitioning operations such as sliding-window joins, consider defining partitions with larger granularity, so fewer partitions are involved in the sliding-window join.
RedistBufSize Performance Implications
To avoid the overhead of sending many small messages across the BYNET, buffers are used to batch individual rows during the row redistribution process. RedistBufSize setting determines the size in kilobytes for row redistribution buffers used during load utilities.
The default is 4 (which translates to 4 KB), and each AMP will have one such buffer in memory for each AMP in the configuration during a load job row redistribution.
If a system has relatively few AMPs, a larger redistribution buffer size usually has a positive effect on load performance. However, on larger systems with many AMPs, a large buffer size can consume excessive memory, especially if many load jobs are run concurrently.
Maintain the RedistBufSize at 4 KB for up to 48 nodes with 8 AMPs/node. As the system configuration grows larger, you can compensate by doing one or more of the following:
- Set RedistBufSize smaller in proportion to the increase in the total number of AMPs, that is, send more smaller messages)
- Add more memory to increase the total memory in the system to accommodate the redistribution buffers somewhat larger.
- Increase the amount of free memory available for redistribution buffers by setting FSGCache percent smaller
To help determine if RedistBufSize is too high, see if the minimum available free memory consistently goes below 100 MB during heavy periods of load utility redistribution. Also, check for significant swapping (more than 10/second) during this period. If this is the case, reduce RedistBufSize an incremental value lower, for example, from 4 KB to 3 KB.