CS186-L18: Parallel Query Processing
Contents
Intro to Parallelism
Architectures and Software Structures
we will focus on the shared-nothing here 😋
Kinds of Query Parallelism
side note:
- intra: single
- inter: multiple at the same level
Parallel Data Acceess
Data Partitioning across Machines
Round robin means that each machine haves the same shuffled data
parallel scans
scan and merge
$\sigma_p$ : an operator that skip entire sites that have no matching tuples in range or hash partitioning
lookup by key
if data partitioned on function of key, then Route lookup only to the relevant nodes
otherwise, broadcast lookup to all nodes
insert
if on function of key, insert only to the relevant nodes
else insert to any nodes
insert an unique key seems to be same
parallel hash join
naive hash join
grace hash join
Pass one is like hashing above, but do it 2x– once for each relation being joined
Pass two is local grace hash join per node
sort-merge join
回到均分问题了
然后和上面一样读取分配两次for join
parallel aggregation/grouping
naive group by:
Symmetric Hash Joins
sort and hash can break the pipeline……
one-sided and Broadcast Joins
one-sided joins
one is sorted/hashed
broadcast joins
one is small
