[jira] [Updated] (GIRAPH-1150) Fixing MemoryEstimatorOracle

["Dionysios Logothetis (Jira)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/GIRAPH-1150?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Dionysios Logothetis updated GIRAPH-1150:
-----------------------------------------
    Fix Version/s: 1.3.0

> Fixing MemoryEstimatorOracle
> ----------------------------
>
>                 Key: GIRAPH-1150
>                 URL: https://issues.apache.org/jira/browse/GIRAPH-1150
>             Project: Giraph
>          Issue Type: Bug
>            Reporter: Hassan Eslami
>            Assignee: Hassan Eslami
>            Priority: Major
>             Fix For: 1.3.0
>
>
> * Once the OOC is at OFFLAODING state, it may offload data to disk indefinitely. We should call GC manually and get OOC out of OFFLOADING if anything has changed in computation/communication pattern.
> * Fixes a "if" statement which entirely disabled regression for memory estimation.
> * Interleaving of in resetting and calculating memory estimation potentially can cause data race.
> * The superstep count in memory estimator was not coherent throughout the calculation.
> * Sometime the memory estimator's accuracy is not good. We should fall back to a threshold-based scheme relying on a more pessimistic memory usage report (such as the one given by JVM -- memory usage includes garbage data too, but it can be used as a pessimistic estimate, as it is currently used in ThresholdBasedOracle).
> More tuning is needed for a smooth memory estimation mechanism.



--
This message was sent by Atlassian Jira
(v8.3.4#803005)