Details
-
Task
-
Status: Closed (View Workflow)
-
Major
-
Resolution: Done
-
1.4.0
-
None
-
2019-06
Description
Some of the defaults are more appropriate to the machines ~ 10-15 years ago. Time to update...
1) PM join threshold. With ssds it makes sense to take some mem from cache and use it for larger pm joins. I would suggest a min of 1GB. The key is named PmMaxMemorySmallSide.
2) Number of aggregation threads & buckets. For some reason this is 4. Would be good to automate this setting, but if time is short, let's just set them higher. As a rough guideline, make the # of buckets 4 X num threads. Keys of interest are RowAggrThreads and RowAggrBuckets.
3) degree of parallelism. This would also be a good candidate for automation. MaxOutstandingRequests is how many extents worth of jobs can be in flight at once. With higher core count, bigger mem, and potential for scale-out, should probably be increased. Automation would be a function of core count across all PMs, ProcessorThreadsPerScan, etc.
There also isn't much reason to have large buffers between jobsteps, so I would suggest trimming FifoSizeLargeSide, and FifoSize. Benchmark a little to confirm.