[
https://issues.apache.org/jira/browse/MAPREDUCE-2162?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12965936#action_12965936
]
Joydeep Sen Sarma commented on MAPREDUCE-2162:
----------------------------------------------
here's the reasoning behind capping stddev at mean/3. we speculate if:
* rate < mean - stddev
implies
* 1/rate > 1/(mean - stddev)
implies
* 1/rate > 1/mean + (1/(mean - stddev) - 1/mean)
implies
# projectedTime > meanTime + Delta
where
* Delta = (1/(mean - stddev) - 1/mean)
if
* stddev <= mean/3 // for example
then
* Delta > (1/(mean - mean/3) - 1/mean) ==>
* Delta > 0.5/mean = 0.5 * MeanTime
now our our equation _1_ becomes:
# projectedTime > MeanTime + 0.5*MeanTime
two observations:
* by capping stddev - we have converted the rate check into a meaningful check
on the running time of a task - tasks that run longer than a certain time
(relative to the mean) will be guaranteed to be speculated.
* the Meantime + 0.5*Meantime slack over the mean is same as the heuristic
discussed in the jira where two rules were discussed:
** dont speculate if runningTime <= MeanTime * 0.5
** dont speculate if remainingTime < MeanTime
* if we add these two together - since runningTime + remainingTime ==
projectedTime - this becomes (roughly):
** speculate only if projectedTime > MeanTime + MeanTime*0.5
so the heuristics in the jira are structurally similar to capping the stddev at
mean/3.
as explained earlier - the percentile stuff is actually (approximately) being
done by speculativeCap (no more than 10% of the tasks can be speculated and
tasks are sorted (by latest finish time) before speculating).
> speculative execution does not handle cases where stddev > mean well
> --------------------------------------------------------------------
>
> Key: MAPREDUCE-2162
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-2162
> Project: Hadoop Map/Reduce
> Issue Type: Bug
> Reporter: Joydeep Sen Sarma
> Assignee: Joydeep Sen Sarma
>
> the new speculation code only speculates tasks whose progress rate deviates
> from the mean progress rate of a job by more than some multiple (typically
> 1.0) of stddev. stddev can be larger than mean. which means that if we ever
> get into a situation where this condition holds true - then a task with even
> 0 progress rate will not be speculated.
> it's not clear that this condition is self-correcting. if a job has thousands
> of tasks - then one laggard task, inspite of not being speculated for a long
> time, may not be able to fix the condition of stddev > mean.
> we have seen jobs where tasks have not been speculated for hours and this
> seems one explanation why this may have happened. here's an example job with
> stddev > mean:
> DataStatistics: count is 6, sum is 1.7141054797775723E-8, sumSquares is
> 2.9381575958035014E-16 mean is 2.8568424662959537E-9 std() is
> 6.388093955645905E-9
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
