giraph-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Maxymilian Śmiech (JIRA) <j...@apache.org>
Subject [jira] [Created] (GIRAPH-837) GiraphStats in MasterCompute has wrong values
Date Sun, 02 Feb 2014 15:48:08 GMT
Maxymilian Śmiech created GIRAPH-837:
----------------------------------------

             Summary: GiraphStats in MasterCompute has wrong values
                 Key: GIRAPH-837
                 URL: https://issues.apache.org/jira/browse/GIRAPH-837
             Project: Giraph
          Issue Type: Bug
          Components: bsp
    Affects Versions: 1.1.0
            Reporter: Maxymilian Śmiech
         Attachments: GiraphStatsExample.java

I am trying to implement new feature (small class assignment) -- voting to change computation
"phase" -- using MasterCompute.

Basically when every Vertex "votes to advance to next phase" I need to check conditions like
in:
{code:title=BspServiceMaster.coordinateSuperstep()|borderStyle=solid}
if (masterCompute.isHalted() ||
        (globalStats.getFinishedVertexCount() ==
        globalStats.getVertexCount() &&
        globalStats.getMessageCount() == 0)) {
      globalStats.setHaltComputation(true);
    }
{code}
In my solution I check if all vertices have "voted" and there are no messages to deliver.
If that is true I broadcast next phase number through an aggregator.

Now I have problems obtaining GlobalStats.getMessageCount(). I see that coordinateSuperstep()
is performing those actions (this is example with superstep 15 and 16):
# Call AbstractComputation.compute() on every Vertex with superstep 15
# Call MasterCompute.compute() with superstep 16
# Aggregate GlobalStats from superstep 15
# Check halting conditions after superstep 15
# Copy GlobalStats to GiraphStats
# Go to step 1

I can access {{GiraphStats.getInstance().getSentMessages()}} in {{MasterCompute.compute()}}

I wrote small test with two nodes and two edges:
- superstep 0:
** MasterCompute does nothing
** each vertex sends one message to all its neighbours
- superstep 1:
** MasterCompute reads GiraphStats.getInstance().getSentMessages().getValue() and broadcasts
it through an IntOverwriteAggregator
** each vertex sets its value to that aggregator's current value and halts

After that, every vertex has value of *0*. I believe this is wrong. If I perform actions from
superstep 1 in superstep 2, I got correct vertex value of *2*.

Explanation is simple: {{MasterCompute.compute()}} called in {{superstep}} has {{GiraphStats}}
from {{superstep-2}}. That is because it is called too early -- before {{GlobalStats}} from
{{superstep-1}} are gathered and put in {{GiraphStats}}.


My proposed solution is:
{code:title=BspServiceMaster.coordinateSuperstep()|borderStyle=solid}
GlobalStats globalStats = aggregateWorkerStats(getSuperstep());
/* ... */
updateCounters(globalStats);
/* ... */
doMasterCompute();
/* ... */
if (masterCompute.isHalted() ||
        (globalStats.getFinishedVertexCount() ==
        globalStats.getVertexCount() &&
        globalStats.getMessageCount() == 0)) {
      globalStats.setHaltComputation(true);
    }
/* ... */
{code}

I am new to Giraph. Maybe this beahvior is intended. Maybe it is not trivial to change. But
for me it is not working as expected.



--
This message was sent by Atlassian JIRA
(v6.1.5#6160)

Mime
View raw message