I was reading Carrie's blog on skew sensitivity of TD12 optimizer.
Here at one point she mentioned,
"Instead of assuming each AMP will receive 1.1 millions rows during row redistribution, the Teradata 12 optimizer assumes a worst-case redistribution of 6 million rows for at least one AMP. It uses that worst-case number of rows to adjust the estimated cost of the row redistribution choice."
Assume that TableA PI column is joined with TableB non-PI column. Definitely TableB will be redistributed to all AMP. Now TableB is a very large table.
As per my understanding in v2r6 it would have assumed that each AMP has received those huge rows and processing in each AMP would have been skewed. (performance of all AMP is equivalent to the slowest AMP).
Does that not happen any more in optimizer?
I mean, currently (from TD12 onwards) will the optimizer be able to pin-point the worst case scenario of row redistribution and accordingly will change its method so that the processing is similar in each AMP?
Please help me to clarify my understanding.
what was the conclusion of carrie, in skewd AMP, the join rocessing for all AMP will be similar or join processing for each AMP will be skewed , as i know
'NUPI’s can create irregular distributions, called "skewed distributions". AMPs that have more than an average number or rows will take longer for full table operations than the other AMPs will. Because an operation is not complete until all AMPs have finished, this will cause the operation to finish less quickly due to being underutilized. ' plus as you said performance is equivalent to the slowest AMP.therefore join processing on each AMP will be skewed or the Optimizer as carrie said will make join processing on each AMP similar?