[
https://issues.apache.org/jira/browse/SYSTEMML-552?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Matthias Boehm resolved SYSTEMML-552.
-------------------------------------
Resolution: Fixed
Assignee: Matthias Boehm
Fix Version/s: SystemML 0.10
> Performance features ALS-CG
> ---------------------------
>
> Key: SYSTEMML-552
> URL: https://issues.apache.org/jira/browse/SYSTEMML-552
> Project: SystemML
> Issue Type: Task
> Reporter: Matthias Boehm
> Assignee: Matthias Boehm
> Fix For: SystemML 0.10
>
>
> Over a spectrum of data sizes, ALS-CG does not always perform as good as we
> would expect due to unnecessary overheads. This task captures related
> performance features:
> 1) Cache-conscious sparse wdivmm left/right: For large factors, the approach
> of iterating through non-zeros in W and computing dot products, leads to
> repeated (unnecessary) scans of the factors from main-memory.
> 2) Preparation sparse W = (X!=0) w/ intrinsics: For scalar operations with
> !=0, there is already a special case which is however unnecessarily
> conservative. We should realize this with a plain memcopy of indices and
> memset 1 for values.
> 3) Flop-aware operator selection QuaternaryOp: For large ranks, all
> quaternary operators become really compute-intensive. In these situations,
> our heuristic of choosing ExecType.CP if the operation fits in driver memory
> does not work very well. Hence, we should take the number of floating point
> operations and the local/cluster degree of parallelism into account when
> deciding for the execution type.
> 4) Improved parallel read sparse binary block: Reading sparse binary block
> matrices with clen>bclen requires a global lock on append and final
> sequential sorting of sparse rows. We should use a more fine-grained locking
> scheme and sort sparse rows in parallel.
> 5) Cache-conscious sparse wsloss all patterns: Similar to wdivmm (see 1) but
> less common since only executed once per outer iteration.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
