Next: Join Product Skew [cont'd.]
Up: No Title
Previous: Extent based Sampling
- In these experiments both the relations were skewed implying that a small number of
repeats can cause a blow up in the result.
Algo (10K Join 10) (1K Join 100) (100 Join 1000)
HH 143.6 144.0 144.0
WRange 148.2 148.2 149.4
VP-RR 49.7 85.8 151.0
VP-RS 56.3 94.1 155.0
- For the above results (table 4) the configuration was same as previous.
- Observations:
(i) The VP-RR & VP-PS have significant success. This
is explained by the fact that in the Range and Weighted Range algorithms
the skew in the relation is not enough to cause tuples with '1's
to be sent to more than one processor.
(ii) In the case of *100 Join *1000 the round robin fails to distribute the
1's because there are so few in the building relation . The processor
scheduling algo fails because it wrongly estimates the work requires per virtual processor.
DBMS
1999-03-11