|
|
Thesis: "Analytical
Modeling for Buffer Hit Rate Prediction"
|
|
Abstract (full text
not available)
|
|
Today's database management
systems (DBMSs) require careful database configuration to achieve optimal
performance, while manually tuning the database becomes more and more impractical.
The facility of "no knobs operation" enables the DBMS to automatically
reallocate its resources to maintain acceptable performance in the face
of changing conditions.
The performance of a DBMS is greatly influenced by the effective use of
main memory, especially the buffer area. Buffer pool size tuning is
therefore crucial to achieving good performance for a DBMS. It is however
a time-consuming process to monitor buffer pool hit rate in order to tune
the system.
In this thesis research, we use an analytical modeling approach to predict
the buffer pool hit rate to simplify the process of estimating hit rate.
Since the buffer replacement algorithm determines the buffer hit rate,
the model must represent the algorithm. A variation of GCLOCK algorithm
is currently used by DB2. We develop a Markov Chain model of GCLOCK
to estimate the hit rates of multiple buffer pools. We evaluate the
accuracy of the model's estimates with experiments carried out on DB2 with
the TPC benchmark.
|
|